Fig. 3: Visualization of CNNs predictions.
From: Deep convolutional neural network-based algorithm for muscle biopsy diagnosis
a Sample H&E-stained images of IIM. b Merged images of H&E-stained images and heatmap images created with Grad-cam. Red color indicates CNN focus areas essential for prediction. c Sample H&E-stained images of non-myositis muscle diseases. d Merged images of H&E-stained and heatmap images created with Grad-cam. e Physicians’ test results in myositis. The left (blue) bar shows the average accuracy of the physicians’ diagnosis with AI-focused images, while the right (orange) bar shows the AI-unfocused images. The error bar indicates the standard deviation. f Confusion matrix of results with AI-focused images in IIM. g Confusion matrix of results with AI-unfocused images in IIM. h Physicians’ test results in non-myositis. The left (blue) bar shows the average accuracy of the pathologists’ diagnosis with AI-focused images, and the right (orange) bar shows the AI-unfocused images. The error bar indicates the standard deviation. i Confusion matrix of results with AI-focused images in non-myositis. j Confusion matrix of results with AI-unfocused images in non-myositis. NA indicates no answer.
