Fig. 2

Diagram for the development and validation of the pathological artificial intelligence diagnostic model. In the training stage, a CNN model (ScanNet) was trained with the training patch, and the patch-level classifier was developed. In the validation stage, the WSI was first divided into validation patches and then input into ScanNet. The outputs of the patches were spliced together to obtain heatmaps. The probability weighted value of each subtype was calculated to give the confidence of WSI-level classification. CNN  convolutional neural network, WSI  whole slide image, HGMI  high-grade muscle invasion, HGNMI  high-grade non-muscle invasion, LGNMI  low-grade non-muscle invasion, IA  illegible area, NIA  normal interstitial area, NA  noise area