Fig. 2: Development and validation of DL systems for tumor differentiation grade recognition.

a Annotations of PDA and WDA in WSI. Yellow, region of poorly differentiated adenocarcinoma (PDA). Red, region of well-differentiated adenocarcinoma (WDA). WSI, whole-slide image. Scale bar, 1 mm. b Samples of tiles for other, PDA, and WDA classes. The tumor class consists of PDA and WDA. Three tiles for each class are shown. Size of tiles, 512 × 512 pixels, 250 × 250 μm. c Pie graph of patient cohorts for tumor differentiation grade recognition. PDA and WDA denote the patients with only PDA and WDA tiles in the tumor class, respectively. PDA and WDA denote the patients with both PDA and WDA tiles. There are 436 patients in total. d Scatter plot for the number of tiles in different classes. Left panel, the plot of the tile count for PDA class versus the other class. Right panel, the plot of the tile count for the WDA class versus the other class. Each circle denotes a patient. Gray circles denote the training dataset. Red circles denote the testing dataset. e DL models to achieve tumor diagnosis. Left panel, the input image of an HE staining tile. Right panel, two DL models for tumor diagnosis, two-class (other/tumor) classification model, and three-class (other/PDA/WDA) classification model. The sample HE staining tile belongs to the tumor class and PDA class. f Heatmap of a confusion matrix for other/tumor classification. The number denotes the count of corresponding tiles. g Heatmap of a confusion matrix for other/PDA/WDA classification. h Gradient-weighted class activation map for PDA. Upper, original HE staining tiles for PDA class. Down, class activation maps for the HE staining tiles. Hot regions correspond to key features for DL models to recognize PDA. The red and blue colors indicate greater importance and less importance, respectively. i Gradient-weighted class activation map for WDA. Hot regions correspond to key features for DL models to recognize WDA.