Figure 4 | Scientific Reports

Figure 4

From: Multiple instance learning detects peripheral arterial disease from high-resolution color fundus photography

Figure 4

Subfigure (a) shows a histogram of all attention weights from the images in the validation set across all k-folds. It indicates that the network ignores many less relevant patches in the input images. All values above 3.5 are combined into the rightmost bar. Attention maps in Subfigure (b) correspond to four correct predictions of PAD. The first two images are true positives and the last two true negatives. The maps suggest that the network mostly relied on patches showing retinal vessels for the prediction.

Back to article page