Skip to main content

Radiomics using computed tomography to predict CD73 expression and prognosis of colorectal cancer liver metastases



Finding a noninvasive radiomic surrogate of tumor immune features could help identify patients more likely to respond to novel immune checkpoint inhibitors. Particularly, CD73 is an ectonucleotidase that catalyzes the breakdown of extracellular AMP into immunosuppressive adenosine, which can be blocked by therapeutic antibodies. High CD73 expression in colorectal cancer liver metastasis (CRLM) resected with curative intent is associated with early recurrence and shorter patient survival. The aim of this study was hence to evaluate whether machine learning analysis of preoperative liver CT-scan could estimate high vs low CD73 expression in CRLM and whether such radiomic score would have a prognostic significance.


We trained an Attentive Interpretable Tabular Learning (TabNet) model to predict, from preoperative CT images, stratified expression levels of CD73 (CD73High vs. CD73Low) assessed by immunofluorescence (IF) on tissue microarrays. Radiomic features were extracted from 160 segmented CRLM of 122 patients with matched IF data, preprocessed and used to train the predictive model. We applied a five-fold cross-validation and validated the performance on a hold-out test set.


TabNet provided areas under the receiver operating characteristic curve of 0.95 (95% CI 0.87 to 1.0) and 0.79 (0.65 to 0.92) on the training and hold-out test sets respectively, and outperformed other machine learning models. The TabNet-derived score, termed rad-CD73, was positively correlated with CD73 histological expression in matched CRLM (Spearman’s ρ = 0.6004; P < 0.0001). The median time to recurrence (TTR) and disease-specific survival (DSS) after CRLM resection in rad-CD73High vs rad-CD73Low patients was 13.0 vs 23.6 months (P = 0.0098) and 53.4 vs 126.0 months (P = 0.0222), respectively. The prognostic value of rad-CD73 was independent of the standard clinical risk score, for both TTR (HR = 2.11, 95% CI 1.30 to 3.45, P < 0.005) and DSS (HR = 1.88, 95% CI 1.11 to 3.18, P = 0.020).


Our findings reveal promising results for non-invasive CT-scan-based prediction of CD73 expression in CRLM and warrant further validation as to whether rad-CD73 could assist oncologists as a biomarker of prognosis and response to immunotherapies targeting the adenosine pathway.


Liver metastases is the most common site of colorectal cancer progression, a malignancy that remains within the top leading causes of cancer-related deaths [1]. Complete resection of colorectal liver metastases (CRLM) combined with systemic chemotherapy, has a curative potential, but approximately 80% of patients recur [2]. Currently, no biomarkers can help identify patients at high risk of early recurrence after CRLM resection, for whom surgery may be futile, who may benefit from adjuvant chemotherapy [3], and who should be followed up more closely based on the risk of recurrence. Consequently, a significant number of patients are burdened by potential complications, side-effects and toxicities associated with treatments not informed by expected prognosis, without overall survival benefits.

Accumulating data support that the immune features of CRLM may be more significantly associated with recurrence and survival after resection, independently from clinical, pathological and tumor genomic features [4]. Extracellular adenosine in the tumor microenvironment (TME) appears as an immunosuppressive mechanism of particular relevance to CRLM microenvironment. The Ecto-5′-nucleotidase, or CD73, expressed by cancer and stromal cells within CRLM, is a rate-limiting enzyme that enhances the breakdown of ATP-derived extracellular adenosine monophosphate into immunosupressive adenosine [5]. The latter binds to the A2A and A2B receptors expressed by T cells, Natural Killer cells, and other immune cells and inhibits their anti-tumoral cytotoxic functions [6]. High intratumoral CD73 expression is strongly associated with poor prognosis in both primary CRC [7] and CRLM [8].

The current histologic assessment of CD73 expression in CRLM, like other histologically-derived biomarkers, is however obtained only after resection of CRLM. Histological-based markers are also not ideal to asses pre-operatively via biopsy given their intratumoral heterogeneous expression [9]. The field still lacks noninvasive CRLM biomarkers of immune features to guide prognostication, which may also, at term, help identify patients most likely to benefit from immunotherapy. Features of computed tomography (CT) images are widely available and part of routine clinical practice, both at baseline and on follow-up examinations for assessment of treatment response, and radiomics biomarkers obtained from these images could help clinical decisions early in the treatment course. Current developments in artificial intelligence span almost all conventional medical image analysis tasks, including the detection and segmentation of anatomical structures, classification and registration of medical images [10]. Radiomics, consisting in extracting quantifiable features from medical images, could detect not only macroscopic characteristics, but also hidden genomic and proteomic properties involved in biological processes [11, 12]. Radiomics studies have shown promising results in the diagnosis of liver diseases, including benign diseases and primary and secondary malignancies, cancer staging and grading and the prediction of patient clinical outcomes such as response to therapy [13, 14]. Concerning the liver, contrast-enhanced CT-scan features have been shown able to detect non-alcoholic steatohepatitis [15], to identify the histological growth pattern of CRLM [16], and to predict response to FOLFOX-based chemotherapy in untreated CRLM patients [17]. To our knowledge, however, CT-scan image features have not been associated with immune features in CRLM [18].

The goal of this study was to test whether we could train a deep learning model with radiomic features extracted from preoperative CT images, to predict postoperative CD73 expression in resected CRLM, and to test whether a probabilistic score predicted by the deep learning model, termed rad-CD73, was associated with patient oncological prognosis.


Study population

This study received the approval of the Institutional Review Board (No. 19.185 for patient consent for biobanking and database; No. 18.023 for project specific with radiomics). We retrospectively studied a cohort of 215 patients who underwent complete resection of CRLM at the Centre Hospitalier de l’Université de Montréal between 2011 and 2014 and were prospectively followed beyond recurrence and until death for clinicopathological and imaging data in a registry. Clinical annotations included demographics, dates and types of all treatments received, and the Clinical Risk Score, calculated by the addition of one point for each of the following features: disease-free interval between the diagnosis of primary tumor and liver metastases of < 12 months; number of metastases > 1; pre-operative carcinoembryonic antigen (CEA) level > 200 ng/mL; largest metastasis > 5 cm; and lymph node positive primary tumor [19]. A liver pathologist reviewed hematoxylin and eosin whole slides of all cases to assess resection margins, the presence of tertiary lymphoid structures, the degree of necrosis and pathological response to preoperative chemotherapy [20].

The inclusion criteria were as follows: (1) CRLM confirmed by histopathological analysis; (2) complete resection of CRLM with curative intent; (3) preoperative, intravenous contrast-enhanced abdominal CT scan available; and (4) CD73 immunofluorescent quantification staining performed. Patients were excluded from the analysis if: (1) preoperative CT images were insufficient to perform feature analysis; (2) CRLM was not visible on the preoperative CT scan or was calcified; (3) histopathological results could not be associated with CT image of a given CRLM on the basis of its description in pathology report and review of the CT-scan. After applying the exclusion criteria, 160 CRLM resected in 122 patients were available for further analysis.

Evaluation of CD73 expression

CD73 expression in CRLM quantified by immunofluorescence were previously generated in this cohort [8]. Briefly, we built tissue microarrays (TMA) using six 0.6 µm TMA cores per CRLM, with up to three CRLM per patient, using FFPE blocks after hematoxylin and eosin review of viable tumor areas by a pathologist and trained resident. We optimized a multiplex immunofluorescence panel to concurrently detect CD73, cytokeratins to compartmentalize stromal and cancer cell expression patterns, and DAPI for nuclear staining of viable cells. Standard deparaffinization and rehydratation protocols were used, followed by antigen retrieval (Dako S1699) in sub-boiling conditions for 40 min, and protein-block (Dako X0909), specific staining with primary antibodies against CD73 (Abcam ab91086, 1:300 dilution) and cytokeratins 8/18 (Dako IR094, 1:2 dilution). We used an anti-mouse IgG1 Alexa-Fluor 647 (Life technology, A21240; 1/800) and anti-rabbit Alexa-Fluor 488 (Life technology, A21206; 1/400) as secondary antibodies, DAPI, and mounted the slides with ProLong Gold (ThermoFisher). Slides were digitalized at 20 × with NanoZoomer-XR (Hamamatsu) and core images imported with TMA maps and identifiers into Visiomorph v.6 software (Visiopharm) for automated quantification. For each core, the percent surface area containing CD73+ cells (expression above background) over the surface area containing all viable cells was calculated, as well as the mean fluorescence intensity in each core. For each CRLM, a mean value was calculated from its corresponding six cores. Patients were stratified into two classes (CD73High and CD73Low) using the median CD73+ positive area score across all CRLM evaluated in this cohort as a cutoff value (3.8%). In patients with more than one CRLM, the mean CD73 was used to classify patients as low or high.

Image preprocessing and radiomics workflow

We analyzed the last contrast-enhanced CT images obtained prior to surgery for CRLM resection and acquired in portal venous phase. The images had a cross-sectional volume size of 512 \(\times\) 512, a mean in-plane resolution of 0.72 mm2 (range = [0.56, 0.98] mm2) and a mean in-depth resolution of 2.26 mm (range = [0.80, 5.0] mm). Images were resampled to isotropic resolution (1 × 1 × 1 mm3) so as to obtain a uniform pixel spacing within the dataset. An automated segmentation algorithm [21] was used to segment CRLM lesions. The segmentation model consists of two convolutional neural networks trained end-to-end in order to jointly segment the liver and the lesions within. A manual examination and refinement of the ensuing 3D segmentations was subsequently carried out by liver radiologists to verify the quality of the volumes of interest to be included in the analysis.

In a subsequent step, radiomic features were extracted from the volumes of interest using PyRadiomics v3.0.1 toolbox [22]. From each image/tumor mask pair, we extracted 107 radiomics features consisting of 18 first-order statistics, 14 shape features and 75 textural features. First-order statistics are histogram-derived features characterizing the distribution of voxel values within the tumor. Shape features encode the 3D shape and size of the region of interest. These features are calculated from approximated shapes, inferred using triangle meshes generated from binary masks using the marching cube technique. Details on the marching cubes algorithm used to build meshes are presented in the Pyradiomics documentation [22]. Thirdly, textural features are derived from predefined matrices and aim to construe the spatial arrangement of voxel intensity values within the lesion. The textural matrices included are the gray-level co-occurrence matrix (GLCM), the gray level dependence matrix (GLDM), the gray-level run length matrix (GLRLM), the gray-level size zone matrix (GLSZM) and the neighboring gray tone difference matrix (NGTDM).

The resulting feature set was standardized for the sake of obtaining a null mean and a unit standard deviation across instances. The least absolute shrinkage and selection operator (LASSO) [23] was used to select the most salient features. The rationale behind this step is to apply an initial coarse dimensionality reduction in order to discard irrelevant features that would otherwise introduce noise into the training process and mislead the model. The LASSO λ hyperparameter (λ = 0.249) was determined by applying five-fold cross validation using the mean squared error as an objective function. Each lesion segmented on CT was matched to the corresponding CRLM in the constructed TMA, using their pathology report block number.

TabNet training and evaluation

In this study, we trained an Attentive Interpretable Tabular Learning (TabNet) model [24] to predict the stratified CD73 expression levels (CD73High vs. CD73Low). TabNet is a deep learning model that incorporates multiple stages of attention modules within its architecture (Additional file 1: Fig. S1). The attention mechanism is based on two transformer blocks: an attentive transformer and a feature transformer. The attentive transformer allows to compute learnable masks, which are used to select the most salient set of features at each step. It also incorporates a prior scale which encodes the degree to which features have been used in the previous steps. The feature transformer processes the filtered features through shared and decision step-specific layers. The outputs of the different decision steps are linearly combined to form the model’s encoder output. Finally, a fully connected layer processes the encoder’s output to obtain the overall output of the model. In order to compare TabNet’s performance with other machine learning models, we also trained an XGBoost, a random forest (RF), a support vector machine (SVM) with a linear kernel function, and a logistic regression (LR) model to perform the same task.

For TabNet and the other baseline models, we first divided the dataset into a training set (125 lesions) and a hold-out test set (35 lesions) for independent validation using data from our center. We then performed a five-fold cross-validation on the first subset. The dataset was hence divided into five separate folds on five iterations and in each iteration, four folds were used for training and the fifth was used for validation (never seen at training). Once cross-validation was completed, an external validation was performed on the hold-out test set. All splits were applied randomly and on a patient-level. In other words, lesions belonging to the same patient were assigned to the same subset in order to ensure that no overlap or information leakage occur between the training and the testing sets. TabNet was trained for 100 epochs on an NVIDIA GeForce GTX TITAN Xp 12 GB with a batch size of 64 and a binary cross-entropy loss function. Adam optimizer [25] was used with a learning rate of 0.02 decaying by 10% after 50 epochs. The area under the receiver operating characteristic curve (AUC) was used to compare the performance of the models.

Model interpretability analysis

Two types of interpretability analyses were carried out in this study: global and local interpretability. In the former case, we attempted to holistically describe the model’s behavior by running a subsequent analysis on the model’s predictions. To this end, we adopted the Shapley Additive Explanations (SHAP) technique [26]. SHAP is an additive feature attribution algorithm that intends to compute, for every feature, a Shapley value which mirrors the contribution of that feature to the model’s final predictions. In this work, we utilized the Kernel SHAP method, which approximates the Shapley values as being the coefficients of a weighted linear regression model, built from a set of sample coalitions. After computing an average Shapley value for each feature over all instances, the features were ranked according to their average Shapley values. Additionally, we made use of the inherent interpretability of the TabNet architecture to provide instance-level explanations of the model’s predictions. To do so, visualizations of the selected feature masks were generated and analyzed, in order to identify the most salient ones for each instance used by the model. The visualization of TabNet’s selected feature importance served two main purposes: (1) to acquire a local explanation of the model’s predictions for each instance and (2) to examine whether the predictive behavior projected by the model’s architecture, was concordant with SHAP results.

Statistical analysis

We used the Wilcoxon rank sum test and Fisher’s exact test for numerical and categorical variables, respectively. Spearman’s correlation coefficient was applied to evaluate correlations. The decision curve analysis [27] was performed using the Python dcurves v0.0.3 package by plotting the net benefit of the predicted biomarker for different threshold probability values [range = (0–1)], given as the minimum probability for which additional testing is recommended. The line representing the “all lesions are CD73Low” hypothesis and the curve representing the “all lesions are CD73High” hypotheses were also plotted for comparison purposes. Survival curves were generated with the Kaplan–Meier method and compared with the log-rank test. Disease-specific survival, for which deaths due to causes other than cancer progression were censored observations, and time to recurrence, were computed from the date of the first hepatectomy. Patients with missing data on the first surgery were excluded from the survival analysis. For survival analysis purposes, TabNet class probabilities, termed rad-CD73, were derived and stratified into rad-CD73High and rad-CD73Low by setting a cut-off value equal to the lower tertile (rad-CD73 > 0.362) based on the distribution of the rad-CD73 score (Additional file 1: Fig. S2), which was consistent with the optimal p-value cut-off proposed by the X-tile software (0.383). For patients with multiple lesions, predicted rad-CD73 scores were averaged across all lesions. Univariate and multivariate Cox proportional hazards regression models were used to generate hazards ratios (HR) with 95% confidence intervals. A two-sided p-value < 0.05 was considered statistically significant. Statistical analyses were conducted using Python Scipy v1.5.3, Python Lifelines v0.27.1 and R Survival v3.4 packages.


The general workflow of the planned analysis is depicted in Fig. 1.

Fig. 1
figure 1

General workflow. Isotropic spatial resampling was applied on preoperative CT-scan images to segment 160 colorectal liver metastases (CRLM) resected in 122 colorectal cancer (CRC) patients who underwent partial hepatectomy. Matching CRLM, identified by their pathology report block numbers and anatomical description, were included in tissue microarrays for automated quantification of CD73 intra-tumoral expression by immunofluorescence. Radiomic features were extracted from the resulting three-dimensional regions of interest (ROI) (18 first-order statistics, 14 shape features and 75 textural features) and preprocessed. Subsequently, an Attentive Interpretable Tabular Learning (TabNet) model was trained to predict CD73 expression dichotomized as CD73High vs. CD73Low based on the surface area expressing CD73 over the total surface of assessable tissue. The model was then evaluated, and its predictions were interpreted using ROC curve and the Shapley Additive Explanations technique (SHAP). The association between radiomic CD73 (rad-CD73) and oncological survival outcomes was analyzed by Kaplan–Meier and logrank test

Patient cohort

The clinicopathological characteristics of the 122 patients treated for resectable CRLM are summarized in Table 1. Mean patient age was 63.4 years (35 to 84), and male patients predominated (63.9%). Preoperative chemotherapy was administered in 78.7% of the patients, consisting of four to six cycles of folfox-based regimen for the vast majority of patients (not shown). Most patients (81.1%) also received chemotherapy after resection of CRLM. Approximately half of the patients were treated for multiple metastasis with a mean number of two (range 1 to 10) and a mean diameter of 4.1 cm (range 1.0 to 20 cm). Based on the composite Clinical Risk Score [19], 38.5% of patients were at high risk of recurrence and death from cancer progression. At time of analysis, the median follow-up was of 57.0 months, during which time 76.2% of patients had recurred and 64.8% died of disease progression.

Table 1 Patient clinicopathological characteristics and correlation with rad-CD73

Prediction of CD73 expression from preoperative CT images

In the training cohort, the ability of the TabNet model to classify CD73High vs. CD73Low lesions was shown to have an AUC of 0.95 (95% confidence interval: 0.87- 1.0). The accuracy, sensitivity and specificity were 0.85, 0.91 and 0.79, respectively. Table 2 summarizes the performance of different models on the hold-out test set. Moreover, TabNet exhibited a high predictive performance on the hold-out test set with an AUC of 0.79 (0.65–0.92). The test set accuracy, sensitivity and specificity were 0.71, 0.63 and 0.79, respectively. Figure 2 depicts the receiver operating characteristic curve and the confusion matrix of the model on the hold-out test set. The model exhibited balanced sensitivity and specificity values, and no class bias was observed. This is reflected by the distribution of the true positives and true negatives in the matrix. We also compared TabNet generalization capability with other machine learning models, which outperformed XGBoost, RF, SVM and LR models. Even though the XGBoost model achieved a satisfactory AUC of 0.61 (0.45–0.77), TabNet outclassed it with a high margin.

Table 2 Performance of the different models on the hold-out test set
Fig. 2
figure 2

Performance of TabNet on the hold-out test set. A Comparison of the ROC curves of TabNet with the other baseline models. TabNet outperformed all the trained baseline models as mirrored by the area under the ROC curves (AUC). B Confusion matrix of TabNet for CD73 classification. LR, Logistic Regression; RF, Random Forest; SVM, Support Vector Machine; TabNet, Attentive Interpretable Tabular Learning; XGBoost, Extreme Gradient Boosting

A significant difference was observed in the predicted TabNet rad-CD73 scores between CD73High and CD73Low lesions (Wilcoxon signed rank test, P < 0.0001) (Fig. 3A). Rad-CD73 scores were positively correlated with CD73 histological expression measured by IF (Spearman’s ρ = 0.600, P < 0.0001) (Fig. 3B). With the goal of assessing the clinical utility of the predicted rad-CD73 score, we performed a decision curve analysis and compared the net benefit of using rad-CD73 with the “treat all as CD73High” and the “treat none as CD73High” strategies. For probability thresholds higher than 0.08, rad-CD73 had a higher net benefit than both the “treat all” and “treat none” approaches (Fig. 4).

Fig. 3
figure 3

Assessment of the predicted TabNet probabilistic score, rad-CD73. A Violin plot depicting the distribution of rad-CD73 score by CD73 expression level. A statistically significant difference was observed in the radiomic score between the CD73High and CD73Low groups (Wilcoxon signed rank test; P < 0.0001). B Spearman’s correlation between rad-CD73 and the actual CD73 expression (the percentage of CD73 positive area per metastasis)

Fig. 4
figure 4

Decision curve analysis. The red line shows the treat none as CD73High approach while the blue line shows the treat all as CD73High approach. For threshold values greater than 0.08, rad-CD73 had a higher net benefit than both the “treat all” and “treat none” approaches

Interpretability of the predictive model

Figure 5 shows representative CT-scan images and corresponding histological images and CD73 IF expression of two CRLM cases with high and low CD73 expression, respectively. Case 1 (left) represents a CRLM with a high CD73 expression (high red IF signal, % CD73 + surface area = 19.24). Concordantly, the corresponding rad-CD73 probabilistic score was 0.69. On the other hand, a low rad-CD73 score (0.06) was attributed to the CRLM of case 2 (right) having a low CD73 IF expression (% CD73 + surface area = 0.37). This finding supported that different CT-scan features could be observed between CD73High and CD73Low CRLM, with homogeneity in CT tumor segmentation in case 1 and heterogeneity in the CT tumor segmentation in case 2. To improve the interpretation of TabNet’s predictions, we applied the SHAP technique and studied the features that contributed the most to the model’s outputs. Figure 6 shows that distinct features had different impact on the model’s output, mirrored by its average Shapley value. Amongst the top five features selected by SHAP, four were textural, reflecting the importance of texture-related characteristics of the lesions in predicting CD73 expression level.

Fig. 5
figure 5

Representative cases with high and low CD73 expression. A A preoperative CT image of a CD73High CRLM (arrow) in the right hemiliver. B A preoperative CT image of a CD73Low CRLM (arrow) in the left hemiliver. C, D Representative tissue microarray cores of the high C and low D CD73 expression of the respective CRLM. Hematoxylin and eosin staining (left) shows the integrity and general architecture of the tissue. Immunofluorescent staining (right) shows cell nucleus (blue, DAPI), CD73 expression (red), and the cancer cells with the pan-cytokeratins epithelial cell marker (green). Scale bars represent 100 µm

Fig. 6
figure 6

TabNet global interpretability analysis using the Shapley Additive Explanations (SHAP) technique. A Summary plot listing radiomics features from top to bottom in a decreasing order of their impact on the model’s decision. The top four most significant features were texture-related features. B Variation of the SHAP value of four selected features with respect to the actual feature value. Homogeneous textures were associated with a higher CD73 expression, as mirrored by the effect of the top-ranking features

The feature with the highest Shapley value was the “Dependence Non Uniformity Normalized” (DNUN), a textural feature computed using the GLDM matrix. The DNUN encodes the heterogeneity in terms of the dependence throughout the lesion: a high DNUN reveals that the image contains regions with disparate dependence levels. The dependence is a term reflecting whether the gray level of a given voxel is dependent on those of the neighboring ones. A region with a low dependence is hence formed of voxels with comparable gray levels whereas the voxels of a high dependence region exhibit discrepancies in their gray levels. SHAP results show that a low DNUN value had a positive impact on the model’s output, prompting an increase in the predicted rad-CD73 score.

Similarly, the textural feature “Size Zone Non Uniformity Normalized” (SZNUN), which encodes the heterogeneity in the size zone volumes of a lesion defined as areas with a constant gray level, had an effect comparable the DNUN on TabNet predictions. The “Small Area Emphasis” (SAE) and the “Informational Measure of Correlation 1” (IMC1) came in third and fourth, respectively. The SAE reflects the prevalence of small zones while the IMC1 encodes the complexity of the textures. Both features had a positive impact on the model’s output; a higher feature value was corresponded to a higher Shapley value. These findings show that the model had the tendency to output high rad-CD73 scores for lesions exhibiting finer textures.

Interestingly, the shape of the lesions also had an impact, albeit less prominent, on the model’s behavior. In fact, SHAP results reveal that spherical lesions were associated with high CD73 expression levels. Finally, Fig. 7 shows that the DNUN was selected for almost all test set instances among the most salient features in TabNet’s third feature selection stage. This finding was consistent with SHAP, as the DNUN had the highest average Shapley value.

Fig. 7
figure 7

Visualization of TabNet learnable mask from the third decision step. The dependence non uniformity normalized (DNUN) feature, designated as the top 1 feature in SHAP analysis, was selected by the learnable mask among the most salient features for almost all test set instances. Brighter colors indicate higher feature importance

Association of rad-CD73 with clinicopathological characteristics and oncological outcomes

We then analyzed the potential clinical significance of rad-CD73 high vs low status. As shown in Table 1, the proportions of rad-CD73High and rad-CD73Low patients were generally similar according to most clinicopathological characteristics, including CRLM size and number, having received chemotherapy or not prior to liver resection, and pathological response to pre-operative chemotherapy. There was hence no statistically significant difference in CRLM diameter (P = 0.830), whether the primary tumor is located in the right or left colon (P = 0.140), the KRAS mutation status (P = 0.735) and the CEA level (P = 0.287) between rad-CD73High and rad-CD73Low CRLM. More patients with rad-CD73High CRLM were however found in those who were diagnosed with liver metastases less than 12 months after the diagnosis of primary CRC (P = 0.037). Although this criterion is one of those constituting the composite clinical risk score (CRS), there were less rad-CD73High patients in those classified as higher risk of recurrence (CRS score 3, 4 or 5).

By univariate analysis, rad-CD73 high vs low status was significantly associated with TTR and DSS, as well as expected clinicopathological features such as the primary tumor depth of invasion (pT stage), high pre-operative CEA, and the composite CRS (Table 3). In the initial immunohistochemical study of CD73 expression in this patient cohort [8], patients with high intratumoral CD73 expression had significantly shorter median TTR compared to low CD73 (11.0 vs 46.4 months) and DSS (19.0 vs. 61.5 months), independent of conventional clinicopathological variables by multivariate analyses. Consistent with the worse prognosis observed in patients bearing tumor with high intratumoral CD73 expression assessed by immunohistochemistry [8, 28,29,30,31], patients with rad-CD73High CRLM had a shorter median TTR of 13.0 months compared to 23.6 months in rad-CD73Low CRLM patients (P = 0.0098). Consistently, the median DSS of rad-CD73High CRLM patients was 53.4 months compared to 126.0 months in rad-CD73Low CRLM patients (P = 0.0222) (Fig. 8). Consistent with the lack of positive association observed between the CRS and rad-CD73, multivariate modeling supported that the prognostic value of rad-CD73 was independent of the CRS for both TTR and DSS (Table 4).

Table 3 Univariate analyses of outcomes according to clinicopathological variables and rad-CD73
Fig. 8
figure 8

Prognostic value of rad-CD73 in colorectal liver metastases (CRLM). A Disease-specific survival according to rad-CD73 and B time-to-recurrence after the initial complete surgical resection of CRLM. In patients with more than one CRLM, the mean rad-CD73 was used to classify patients as low or high. The lower tertile was used as a cut-off value (rad-CD73 > 0.362)

Table 4 Multivariate analyses of outcomes according to clinical risk score and rad-CD73


In this study, we developed a noninvasive imaging surrogate of CD73 by leveraging state-of-the-art deep learning techniques trained with radiomic features. Despite recent progress in prognostication based on immune features of CRLM resected with curative intent [4], there are no noninvasive immune biomarkers that ultimately may guide clinical decision making. To our knowledge, this is the first work developing and testing a machine learning tool to predict immunosuppressive CD73 expression from diagnostic CT images. The proposed model exhibited good performance in classifying CRLM lesions into CD73High and CD73Low groups. We also demonstrated that the predicted rad-CD73 score was highly correlated with the actual expression as measured in vitro by immunohistochemistry. Moreover, the clinical significance of rad-CD73 was supported by its association with patient prognosis.

Radiomics has achieved major breakthroughs in recent years in precision oncology and has paved the way for individualized patient care. Its clinical applications include disease diagnosis, prognosis and treatment planning [32, 33]. It could hence be applied for cancer detection, allowing for a noninvasive differentiation between benign and malignant neoplasms and consequently minimizing the unwarranted collection of tissue samples. Moreover, radiomics was coupled with conventional diagnostic tools in order to augment their sensitivity to detect diseases early in their development [34]. Finally, it was shown to be able to forecast oncological outcomes of patients such as survival, recurrence, response to adjuvant therapy and metastatic progression [35]. The main advantage of radiomics over conventional techniques is that it provides a holistic and noninvasive assessment of the tissues, as opposed to more invasive histopathological tissue analysis methods and RNA sequencing which require biopsies taken from tumor regions. Its performance is hence less affected by the tumoral heterogeneity associated with the biopsy site. Moreover, it is less labor intensive and allows a quicker profiling of the patients than existing diagnostic tools. Nevertheless, the translation of radiomics to the clinic has been hindered by some challenges including the current retrospective design of the majority of the radiomics studies and the “black box” nature of the predictive models. Therefore, radiomics should be considered as a promising complementary tool for personalized treatment to be validated prospectively.

In the past few years, several efforts have been made for the interpretation of the results produced by machine learning models, which are still considered as black boxes. Model interpretability is particularly important in the medical oncology field to ensure traceability and informed clinical decision-making. Interpretability techniques could be divided into two major categories: model-specific and model-agnostic [36]. The former interpretability is acquired in models that are conceived to be inherently explainable through attention mechanisms for instance, which provides feedback on regions of focus in deep learning [37, 38]. On the other hand, model-agnostic interpretability techniques are usually implemented separately and do not depend on the model’s architecture. While the latter involves the implementation of an additional step, it presents the advantage of being applicable to a variety of models. In this work, we tested both techniques. We first trained a TabNet model and sought to understand its instance-level decisions by visualizing its feature selection masks. In a post-hoc analysis, we applied SHAP technique to decipher its overall behavior. The study revealed that the most salient features for the prediction of CD73 expression were texture-related. Moreover, textural heterogeneity was associated with a lower CD73 expression and inversely applicable. This was mirrored by the impact of the DNUN, the SZNUN and the IMC1 on the model’s predictions. These findings are in line with several previous studies associating textural features with response to immunotherapy [39]. Tang et al. [40] found a cluster of non-small cell lung cancer patients exhibiting concurrent tumoral heterogeneity and high CD3 T cell infiltration. Yoon et al. [41] showed that type 2 helper T cells were associated with high variance and IMC, mirroring lesion heterogeneity. On the other hand, Sun et al. [42] found, in a study combining several cancer sites, that a high CD8 score, indicative of inflammatory infiltrate, was associated with homogeneous lesion appearances. They also attributed heterogeneity in pixel intensities to convoluted underlying processes such as excessive disorderly tumor vascularization. CT-scan image features may also be enhanced by the use of nanoparticles, characterized by a high permeability and retention in tumors. Several nanoparticles have been tested [43], including silver nanoparticles [44]. In this context, Devkota et al. [45] demonstrated that radiomics features, extracted from nanoparticle contrast-enhanced CT rather than conventional imaging, were better suited for the prediction of response to cellular immunotherapy.

Immunotherapy has become a mainstay in the treatment of several advanced malignancies. This has motivated several researchers to leverage radiomics to forecast response to immunotherapeutic agents [46,47,48] either by predicting established biomarkers within the tumor microenvironment or by attempting to directly associate imaging features with patient outcomes, such as radiological response, survival and time-to-recurrence. However, the vast majority of the conducted studies focus on non-small cell lung cancer given the universal availability of chest images and the proven effectiveness of immune checkpoint inhibitors (ICI) in advanced lung cancer. The application of radiomics in CRLM immuno-oncology remains vastly unexplored. While prior work on radiomics of CRLM have focused on response to chemotherapy as measured by the tumor regression grade [49] or the RECIST criteria [50, 51], and tumor histological features such as the histological growth patterns [16], we aimed to develop and validate a radiomic immune marker for CRLM.

While first generation ICI such as anti-programmed cell death protein 1 (PD1), anti-PD1 ligand and anti-cytotoxic T lymphocyte antigen 4 have proven to be effective in some cancers [52], 95% of metastatic CRC are refractory to these immunotherapies [53, 54]. The underlying mechanisms driving these poor outcomes include tumor heterogeneity and the coexistence of complex immune escape mechanisms within the hepatic tumor microenvironment. Because of the immunosuppressive nature of the adenosine pathway, adenosinergic molecules are now being explored for the development of novel therapeutic agents [55]. In particular, CD73 ectonucleotidase plays a major role in the generation of immunosuppressive adenosine and has recently emerged as a novel immunotherapeutic target that can be blocked by monoclonal antibodies, while adenosine receptor inhibitors are also being tested in early phase trials [56, 57].

Overall, the clinical use of high rad-CD73 on CT-scan imaging of patients with resectable CRLM, as it may identify a subset of patients with earlier recurrence, death, and higher intratumoral CD73 expression, could be tested prospectively in many ways to determine whether: a) closer follow-up after CRLM resection could lead to earlier treatment of recurrence and survival benefits; b) adjuvant systemic therapy after CRLM resection could reduce the risk of recurrence and death; and c) it can predict the efficacy of patients more likely to respond to anti-CD73 or adenosine receptor inhibitors.

Our work has some limitations notwithstanding. First, because this is a single-center investigation, it will be important to verify the reproducibility of the results by testing the model on a cohort of patients recruited in different institutions. Leveraging epidemiologically diversified databases is of equal importance in order to minimize any bias that could be introduced by unrepresentative datasets. Second, the sample size is relatively limited to deploy deep learning models. Nevertheless, radiomic pipelines have the advantage of being more transparent than end-to-end black box models due to the interpretability of radiomic features. Third, our work does not take into account the dynamic aspect of the tumor microenvironment, with varying delays between the preoperative CT and the histopathological analysis from one patient to another. Future studies should take into account the temporal fluctuations when training artificial intelligence tools to characterize tumor biology. This is markedly true in radiomics-based analyses since medical images could be leveraged in longitudinal studies as a result of their omnipresence.


In this study, we introduced a deep learning pipeline for the prediction of CD73 expression in curatively resected CRLM from preoperative CT images. The conceived rad-CD73 biomarker could serve as a noninvasive, fast and low cost tool to identify candidates for targeted immunotherapy. Due to its association with patient prognosis, it could also be leveraged to assist oncologists to personalize the need for adjuvant treatments and the intensity of follow-up strategies. The generalizability of the model needs to be validated on independent, large and epidemiologically diverse cohorts, and its impact on clinical decision making will need to be tested prospectively.

Availability of data and materials

All raw tissue microarray images and associated marker quantification and patient clinicopathological characteristics are available upon request from the study authors.



Area under the receiver operating characteristic curve


Carcinoembryonic antigen


Colorectal cancer liver metastases


Colorectal cancer


Clinical risk score


Computed tomography


Dependence non uniformity normalized


Gray-level co-occurrence matrix


Gray level dependence matrix


Gray-level run length matrix


Gray-level size zone matrix


Hazards ratios


Immune checkpoint inhibitors


Informational measure of correlation


Least absolute shrinkage and selection operator


Logistic regression


Neighboring gray tone difference matrix


Programmed cell death protein 1


Random forest


Support vector machine


Small area emphasis


Shapley additive explanations


Size zone non uniformity normalized


Attentive interpretable tabular learning


Tissue microarrays


  1. Siegel RL, Miller KD, Goding Sauer A, Fedewa SA, Butterly LF, Anderson JC, et al. Colorectal cancer statistics, 2020. CA Cancer J Clin. 2020;70(3):145–64.

    Article  PubMed  Google Scholar 

  2. Tomlinson JS, Jarnagin WR, DeMatteo RP, Fong Y, Kornprat P, Gonen M, et al. Actual 10-year survival after resection of colorectal liver metastases defines cure. J Clin Oncol. 2007;25(29):4575–80.

    Article  PubMed  Google Scholar 

  3. Kanemitsu Y, Shimizu Y, Mizusawa J, Inaba Y, Hamaguchi T, Shida D, et al. Hepatectomy followed by mFOLFOX6 versus hepatectomy alone for liver-only metastatic colorectal cancer (JCOG0603): a phase II or III randomized controlled trial. J Clin Oncol. 2021;39(34):3789–99.

    Article  CAS  PubMed  Google Scholar 

  4. Baldin P, Van den Eynde M, Mlecnik B, Bindea G, Beniuga G, Carrasco J, et al. Prognostic assessment of resected colorectal liver metastases integrating pathological features, RAS mutation and immunoscore. J Pathol Clin Res. 2021;7(1):27–41.

    Article  CAS  PubMed  Google Scholar 

  5. Zhang B, Song B, Wang X, Chang XS, Pang T, Zhang X, et al. The expression and clinical significance of CD73 molecule in human rectal adenocarcinoma. Tumour Biol. 2015;36(7):5459–66.

    Article  CAS  PubMed  Google Scholar 

  6. Leone RD, Emens LA. Targeting adenosine for cancer immunotherapy. J Immunother Cancer. 2018;6(1):57.

    Article  PubMed  PubMed Central  Google Scholar 

  7. Wu XR, He XS, Chen YF, Yuan RX, Zeng Y, Lian L, et al. High expression of CD73 as a poor prognostic biomarker in human colorectal cancer. J Surg Oncol. 2012;106(2):130–7.

    Article  CAS  PubMed  Google Scholar 

  8. Messaoudi N, Cousineau I, Arslanian E, Henault D, Stephen D, Vandenbroucke-Menu F, et al. Prognostic value of CD73 expression in resected colorectal cancer liver metastasis. Oncoimmunology. 2020;9(1):1746138.

    Article  PubMed  PubMed Central  Google Scholar 

  9. Eide PW, Moosavi SH, Eilertsen IA, Brunsell TH, Langerud J, Berg KCG, et al. Metastatic heterogeneity of the consensus molecular subtypes of colorectal cancer. NPJ Genom Med. 2021;6(1):59.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  10. Litjens G, Kooi T, Bejnordi BE, Setio AAA, Ciompi F, Ghafoorian M, et al. A survey on deep learning in medical image analysis. Med Image Anal. 2017;42:60–88.

    Article  PubMed  Google Scholar 

  11. Limkin EJ, Sun R, Dercle L, Zacharaki EI, Robert C, Reuze S, et al. Promises and challenges for the implementation of computational medical imaging (radiomics) in oncology. Ann Oncol. 2017;28(6):1191–206.

    Article  CAS  PubMed  Google Scholar 

  12. Lambin P, Rios-Velazquez E, Leijenaar R, Carvalho S, van Stiphout RG, Granton P, et al. Radiomics: extracting more information from medical images using advanced feature analysis. Eur J Cancer. 2012;48(4):441–6.

    Article  PubMed  PubMed Central  Google Scholar 

  13. Hu W, Yang H, Xu H, Mao Y. Radiomics based on artificial intelligence in liver diseases: where we are? Gastroenterol Rep. 2020;8(2):90–7.

    Article  Google Scholar 

  14. Nam D, Chapiro J, Paradis V, Seraphin TP, Kather JN. Artificial intelligence in liver diseases: improving diagnostics, prognostics and response prediction. JHEP Rep. 2022;4(4):100443.

    Article  PubMed  PubMed Central  Google Scholar 

  15. Naganawa S, Enooku K, Tateishi R, Akai H, Yasaka K, Shibahara J, et al. Imaging prediction of nonalcoholic steatohepatitis using computed tomography texture analysis. Eur Radiol. 2018;28(7):3050–8.

    Article  PubMed  Google Scholar 

  16. Cheng J, Wei J, Tong T, Sheng W, Zhang Y, Han Y, et al. Prediction of histopathologic growth patterns of colorectal liver metastases with a noninvasive imaging method. Ann Surg Oncol. 2019;26(13):4587–98.

    Article  PubMed  Google Scholar 

  17. Maaref A, Romero FP, Montagnon E, Cerny M, Nguyen B, Vandenbroucke F, et al. Predicting the response to FOLFOX-based chemotherapy regimen from untreated liver metastases on baseline CT: a deep neural network approach. J Digit Imaging. 2020;33(4):937–45.

    Article  PubMed  PubMed Central  Google Scholar 

  18. Fiz F, Vigano L, Gennaro N, Costa G, La Bella L, Boichuk A, et al. Radiomics of liver metastases: a systematic review. Cancers. 2020;12(10):2881.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  19. Fong Y, Fortner J, Sun RL, Brennan MF, Blumgart LH. Clinical score for predicting recurrence after hepatic resection for metastatic colorectal cancer: analysis of 1001 consecutive cases. Ann Surg. 1999;230(3):309–18.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  20. Rubbia-Brandt L, Giostra E, Brezault C, Roth AD, Andres A, Audard V, et al. Importance of histological tumor response assessment in predicting the outcome in patients with colorectal liver metastases treated with neo-adjuvant chemotherapy followed by liver surgery. Ann Oncol. 2007;18(2):299–304.

    Article  CAS  PubMed  Google Scholar 

  21. Vorontsov E, Chartrand G, Tang A, Pal C, Kadoury S. Liver lesion segmentation informed by joint liver segmentation. In: 2018 IEEE 15th International Symposium on Biomedical Imaging (ISBI 2018). 2018. pp 1332–5.

  22. van Griethuysen JJM, Fedorov A, Parmar C, Hosny A, Aucoin N, Narayan V, et al. Computational radiomics system to decode the radiographic phenotype. Cancer Res. 2017;77(21):e104–7.

    Article  PubMed  PubMed Central  Google Scholar 

  23. Tibshirani R. Regression shrinkage and selection via the lasso. J Roy Stat Soc Ser B. 1996;58(1):267–88.

    Google Scholar 

  24. Arik SÖ, Pfister T. TabNet: attentive interpretable tabular learning. arXiv. 2021.

    Article  Google Scholar 

  25. Kingma DP, Ba J. Adam: a method for stochastic optimization. arXiv. 2014.

    Article  Google Scholar 

  26. Lundberg SM, Lee S-I. A unified approach to interpreting model predictions. arXiv. 2017.

    Article  Google Scholar 

  27. Vickers AJ, Elkin EB. Decision curve analysis: a novel method for evaluating prediction models. Med Decis Making. 2006;26(6):565–74.

    Article  PubMed  PubMed Central  Google Scholar 

  28. Buisseret L, Pommey S, Allard B, Garaud S, Bergeron M, Cousineau I, et al. Clinical significance of CD73 in triple-negative breast cancer: multiplex analysis of a phase III clinical trial. Ann Oncol. 2018;29(4):1056–62.

    Article  CAS  PubMed  Google Scholar 

  29. Leclerc BG, Charlebois R, Chouinard G, Allard B, Pommey S, Saad F, et al. CD73 expression is an independent prognostic factor in prostate cancer. Clin Cancer Res. 2016;22(1):158–66.

    Article  CAS  PubMed  Google Scholar 

  30. Jacoberger-Foissac C, Cousineau I, Bareche Y, Allard D, Chrobak P, Allard B, et al. CD73 inhibits cGAS-STING and cooperates with CD39 to promote pancreatic cancer. Cancer Immunol Res. 2023;11(1):56–71.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  31. Turcotte M, Spring K, Pommey S, Chouinard G, Cousineau I, George J, et al. CD73 is associated with poor prognosis in high-grade serous ovarian cancer. Cancer Res. 2015;75(21):4494–503.

    Article  CAS  PubMed  Google Scholar 

  32. Parekh VS, Jacobs MA. Deep learning and radiomics in precision medicine. Expert Rev Precis Med Drug Dev. 2019;4(2):59–72.

    Article  PubMed  PubMed Central  Google Scholar 

  33. Aerts HJ. The potential of radiomic-based phenotyping in precision medicine: a review. JAMA Oncol. 2016;2(12):1636–42.

    Article  PubMed  Google Scholar 

  34. Gillies RJ, Schabath MB. Radiomics improves cancer screening and early detection. Cancer Epidemiol Biomark Prev. 2020;29(12):2556–67.

    Article  CAS  Google Scholar 

  35. Li S, Zhou B. A review of radiomics and genomics applications in cancers: the way towards precision medicine. Radiat Oncol. 2022;17(1):217.

    Article  PubMed  PubMed Central  Google Scholar 

  36. Stiglic G, Kocbek P, Fijacko N, Zitnik M, Verbert K, Cilar L. Interpretability of machine learning-based prediction models in healthcare. WIREs Data Min Knowl Discov. 2020;10(5):e1379.

    Google Scholar 

  37. Vaswani A, Shazeer NM, Parmar N, Uszkoreit J, Jones L, Gomez AN, et al. Attention is all you need. arXiv. 2017.

    Article  Google Scholar 

  38. Mohankumar AK, Nema P, Narasimhan S, Khapra MM, Srinivasan BV, Ravindran B. Towards transparent and explainable attention models. arXiv. 2020.

    Article  Google Scholar 

  39. Wang JH, Wahid KA, van Dijk LV, Farahani K, Thompson RF, Fuller CD. Radiomic biomarkers of tumor immune biology and immunotherapy response. Clin Transl Radiat Oncol. 2021;28:97–115.

    Article  PubMed  PubMed Central  Google Scholar 

  40. Tang C, Hobbs B, Amer A, Li X, Behrens C, Canales JR, et al. Development of an immune-pathology informed radiomics model for non-small cell lung cancer. Sci Rep. 2018;8(1):1922.

    Article  PubMed  PubMed Central  Google Scholar 

  41. Yoon HJ, Kang J, Park H, Sohn I, Lee SH, Lee HY. Deciphering the tumor microenvironment through radiomics in non-small cell lung cancer: correlation with immune profiles. PLoS ONE. 2020;15(4):e0231227.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  42. Sun R, Limkin EJ, Vakalopoulou M, Dercle L, Champiat S, Han SR, et al. A radiomics approach to assess tumour-infiltrating CD8 cells and response to anti-PD-1 or anti-PD-L1 immunotherapy: an imaging biomarker, retrospective multicohort study. Lancet Oncol. 2018;19(9):1180–91.

    Article  CAS  PubMed  Google Scholar 

  43. Eftekhari A, Kryschi C, Pamies D, Gulec S, Ahmadian E, Janas D, et al. Natural and synthetic nanovectors for cancer therapy. Nanotheranostics. 2023;7(3):236–57.

    Article  PubMed  PubMed Central  Google Scholar 

  44. Ramazanli VN, Ahmadov IS. Synthesis of silver nanoparticles by using extract of olive leaves. Adv Biol Earth Sci. 2022;7(3):238–44.

    Google Scholar 

  45. Devkota L, Starosolski Z, Rivas CH, Stupin I, Annapragada A, Ghaghada KB, et al. Detection of response to tumor microenvironment-targeted cellular immunotherapy using nano-radiomics. Sci Adv. 2020;6(28):eaba6156.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  46. Porcu M, Solinas C, Mannelli L, Micheletti G, Lambertini M, Willard-Gallo K, et al. Radiomics and “radi-…omics” in cancer immunotherapy: a guide for clinicians. Crit Rev Oncol Hematol. 2020;154:103068.

    Article  PubMed  Google Scholar 

  47. Kang CY, Duarte SE, Kim HS, Kim E, Park J, Lee AD, et al. Artificial Intelligence-based Radiomics in the Era of Immuno-oncology. Oncologist. 2022;27(6):e471–83.

    Article  PubMed  PubMed Central  Google Scholar 

  48. Bodalal Z, Trebeschi S, Nguyen-Kim TDL, Schats W, Beets-Tan R. Radiogenomics: bridging imaging and genomics. Abdom Radiol. 2019;44(6):1960–84.

    Article  Google Scholar 

  49. Rao S-X, Lambregts DM, Schnerr RS, Beckers RC, Maas M, Albarello F, et al. CT texture analysis in colorectal liver metastases: a better way than size and volume measurements to assess response to chemotherapy? United Eur Gastroenterol J. 2016;4(2):257–63.

    Article  CAS  Google Scholar 

  50. Ahn SJ, Kim JH, Lee SM, Park SJ, Han JK. CT reconstruction algorithms affect histogram and texture analysis: evidence for liver parenchyma, focal solid liver lesions, and renal cysts. Eur Radiol. 2019;29(8):4008–15.

    Article  PubMed  Google Scholar 

  51. Andersen IR, Thorup K, Andersen MB, Olesen R, Mortensen FV, Nielsen DT, et al. Texture in the monitoring of regorafenib therapy in patients with colorectal liver metastases. Acta Radiol. 2019;60(9):1084–93.

    Article  PubMed  Google Scholar 

  52. Johnson DB, Sullivan RJ, Menzies AM. Immune checkpoint inhibitors in challenging populations. Cancer. 2017;123(11):1904–11.

    Article  PubMed  Google Scholar 

  53. Yu X, Zhu L, Liu J, Xie M, Chen J, Li J. Emerging role of immunotherapy for colorectal cancer with liver metastasis. Onco Targets Ther. 2020;13:11645–58.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  54. Dai Y, Zhao W, Yue L, Dai X, Rong D, Wu F, et al. Perspectives on immunotherapy of metastatic colorectal cancer. Front Oncol. 2021;11:659964.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  55. Vijayan D, Young A, Teng MWL, Smyth MJ. Targeting immunosuppressive adenosine in cancer. Nat Rev Cancer. 2017;17(12):709–24.

    Article  CAS  PubMed  Google Scholar 

  56. Zhang B. CD73: a novel target for cancer immunotherapy. Cancer Res. 2010;70(16):6407–11.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  57. Roh M, Wainwright DA, Wu JD, Wan Y, Zhang B. Targeting CD73 to augment cancer immunotherapy. Curr Opin Pharmacol. 2020;53:66–76.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

Download references


The authors would like to thank L. Rousseau, S. Langevin and J. Bilodeau for the recruitment of patients and the maintenance of clinicopathological data; L. Meunier and V. Barès for building the tissue microarray and scanning; A. Cleret-Bohot from the CRCHUM for assisting in the automated quantification of cells and markers; M. Cerny, V. Hamilton, T. Derennes and A. Ilinca for editing segmentations and distinguishing liver metastases from benign coexisting liver lesions.


This work was supported by the Canada Research Chairs, by the National Science and Engineering Research Council of Canada (NSERC) Discovery grant RGPIN-2020-06558 and the Université de Montréal Roger Des Groseillers Research Chair in Hepatopancreatobiliary Surgical Oncology. ST and SK are scientists of the Centre de recherche du Centre hospitalier de l’Université de Montréal (CRCHUM) supported by the Fonds de recherche du Québec—Santé (FRQ-S). ST was supported by the FRQ-S Young Clinician Scientist Seed Grant (No. 32633), the FRQS Clinician Scientist Junior-1&2 Salary Award (No. 30861, No. 298832), and the Institut du Cancer de Montréal establishment award. DH was supported by the FRQ-S phase 1 award for medical resident engaged in clinician-scientist training. NM was supported by the International Hepato-Pancreato-Biliary Association (IHPBA) Kenneth Warren Research Fellowship and Ethicon Inc. (Johnson & Johnson).

Author information

Authors and Affiliations



SK conceived the initial project of predicting immunological profiles from CT images and ST provided domain expertise in CRLM. NM, DH, ST and JS generated CD73 expression data. GS supervised block selection for TMA construction. ST, RS, DH and RR performed CRLM matching. AT and EM supervised CRLM segmentation on CT. RS developed the framework, trained the models, performed statistical analyses and wrote the original paper draft. All authors provided critical revision of the manuscript according to their domain knowledge. All authors read and approved the final manuscript.

Corresponding authors

Correspondence to Simon Turcotte or Samuel Kadoury.

Ethics declarations

Ethics approval and consent to participate

This study was approved by the Centre Hospitalier de l’Université de Montréal (CHUM) institutional review board (No. 16.262) and was performed in accordance with the Declaration of Helsinki. All patients provided informed written consent to participate to the CHUM hepatopancreatobiliary cancer biobank and prospective registry associated with this study (No. 09.237).

Consent for publication

Does not apply.

Competing interests

Outside the submitted work, ST has received consultant fees from Bristol Myers Squibb and Turnstone Biologics, speaking fees from Celgene and Astra Zeneca, and has research funding from Lovance Biotherapeutics and Turnstone Biologics. JS is a permanent member of the scientific advisory board of Surface Oncology and holds stocks of Surface Oncology.

Additional information

Publisher's Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary Information

Additional file 1

: Fig. S1. TabNet model architecture depicts the architecture of the deep learning model used in this work. Fig. S2. Distribution of the predicted TabNet probabilistic score, rad-CD73, across the patients.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit The Creative Commons Public Domain Dedication waiver ( applies to the data made available in this article, unless otherwise stated in a credit line to the data.

Reprints and Permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Saber, R., Henault, D., Messaoudi, N. et al. Radiomics using computed tomography to predict CD73 expression and prognosis of colorectal cancer liver metastases. J Transl Med 21, 507 (2023).

Download citation

  • Received:

  • Accepted:

  • Published:

  • DOI: