Relevance mass accuracy (RMA) and relevance rank accuracy (RRA) on the LRP-α1β0 explanation heatmaps of images of the IDRiD dataset.

The results show that while supervised models overfit on the hard exudates, the self-supervised models look at diverse set of input features (lesions). On the other hand, we also find that IFI models show higher accuracies when compared to IWNE models.

研究结果表明，尽管监督模型在硬性渗出灶（hard exudates）任务上出现过拟合现象，但自监督模型会关注多样化的输入特征（即各类病灶）。另一方面，我们还发现，相较于IWNE模型，IFI模型的准确率更高。