Fig. 2.

Schematic workflow of the 2D and 3D CNN models based on Xception. A For 2D CNN, a single shoulder slide was the input, while 2D convolution layers were utilized to extract image features. Finally, 2048 features were extracted and fed into a classifier, from which the output was the probabilities of tear and normal. B For the 3D CNN model, 3D shoulder image blocks were the input, and 3D convolution layers were utilized to extract image features. Finally, 2048 features were extracted and fed into a classifier, from which the output was the probabilities of tear and normal. CNN, convolutional neural network