Influence of training and expertise on deep neural network attention and human attention during a medical image classification task.

In many different domains, experts can make complex decisions after glancing very briefly at an image. However, the perceptual mechanisms underlying expert performance are still largely unknown. Recently, several machine learning algorithms have been shown to outperform human experts in specific tasks. But these algorithms often behave as black boxes and their information processing pipeline remains unknown. This lack of transparency and interpretability is highly problematic in applications involving human lives, such as health care. One way to "open the black box" is to compute an artificial attention map from the model, which highlights the pixels of the input image that contributed the most to the model decision. In this work, we directly compare human visual attention to machine visual attention when performing the same visual task. We have designed a medical diagnosis task involving the detection of lesions in small bowel endoscopic images. We collected eye movements from novices and gastroenterologist experts while they classified medical images according to their relevance for Crohn's disease diagnosis. We trained three state-of-the-art deep learning models on our carefully labeled dataset. Both humans and machine performed the same task. We extracted artificial attention with six different post hoc methods. We show that the model attention maps are significantly closer to human expert attention maps than to novices', especially for pathological images. As the model gets trained and its performance gets closer to the human experts, the similarity between model and human attention increases. Through the understanding of the similarities between the visual decision-making process of human experts and deep neural networks, we hope to inform both the training of new doctors and the architecture of new algorithms.

Multi-expert annotation of Crohn's disease images of the small bowel for automatic detection using a convolutional recurrent attention neural network.

Background and study aims Computer-aided diagnostic tools using deep neural networks are efficient for detection of lesions in endoscopy but require a huge number of images. The impact of the quality of annotation has not been tested yet. Here we describe a multi-expert annotated dataset of images extracted from capsules from Crohn's disease patients and the impact of the quality of annotations on the accuracy of a recurrent attention neural network. Methods Images of capsule were annotated by a reader first and then reviewed by three experts in inflammatory bowel disease. Concordance analysis between experts was evaluated by Fleiss' kappa and all the discordant images were, again, read by all the endoscopists to obtain a consensus annotation. A recurrent attention neural network developed for the study was tested before and after the consensus annotation. Available neural networks (ResNet and VGGNet) were also tested under the same conditions. Results The final dataset included 3498 images with 2124 non-pathological (60.7 %), 1360 pathological (38.9 %), and 14 (0.4 %) inconclusive. Agreement of the experts was good for distinguishing pathological and non-pathological images with a kappa of 0.79 ( P  < 0.0001). The accuracy of our classifier and the available neural networks increased after the consensus annotation with a precision of 93.7 %, sensitivity of 93 %, and specificity of 95 %. Conclusions The accuracy of the neural network increased with improved annotations, suggesting that the number of images needed for the development of these systems could be diminished using a well-designed dataset.