Vision Transformer(ViT) is one of the most widely used models in the computer vision field with its great performance on various tasks. In order to fully utilize the ViT-based architecture in various applications, proper visualization methods with a decent localization performance are necessary, but these methods employed in CNN-based models are still not available in ViT due to its unique structure. In this work, we propose an attention-guided visualization method applied to ViT that provides a high-level semantic explanation for its decision. Our method selectively aggregates the gradients directly propagated from the classification output to each self-attention, collecting the contribution of image features extracted from each location of the input image. These gradients are additionally guided by the normalized self-attention scores, which are the pairwise patch correlation scores. They are used to supplement the gradients on the patch-level context information efficiently detected by the self-attention mechanism. This approach of our method provides elaborate high-level semantic explanations with great localization performance only with the class labels. As a result, our method outperforms the previous leading explainability methods of ViT in the weakly-supervised localization task and presents great capability in capturing the full instances of the target class object. Meanwhile, our method provides a visualization that faithfully explains the model, which is demonstrated in the perturbation comparison test.

Deep learning has become an essential tool in medical image analysis owing to its remarkable performance. Target classification and model interpretability are key applications of deep learning in medical image analysis, and hence many deep learning-based algorithms have emerged. Many existing deep learning-based algorithms include pooling operations, which are a type of subsampling used to enlarge the receptive field. However, pooling operations degrade the image details in terms of signal processing theory, which is significantly sensitive to small objects in an image. Therefore, in this study, we designed a Rense block and edge conservative module to effectively manipulate previous feature information in the feed-forward learning process. Specifically, a Rense block, an optimal design that incorporates skip connections of residual and dense blocks, was demonstrated through mathematical analysis. Furthermore, we avoid blurring of the features in the pooling operation through a compensation path in the edge conservative module. Two independent CT datasets of kidney stones and lung tumors, in which small lesions are often included in the images, were used to verify the proposed RenseNet. The results of the classification and explanation heatmaps show that the proposed RenseNet provides the best inference and interpretation compared to current state-of-the-art methods. The proposed RenseNet can significantly contribute to efficient diagnosis and treatment because it is effective for small lesions that might be misclassified or misinterpreted.

Vision Transformer(ViT) is one of the most widely used models in the computer vision field with its great performance on various tasks. In order to fully utilize the ViT-based architecture in various applications, proper visualization methods with a decent localization performance are necessary, but these methods employed in CNN-based models are still not available in ViT due to its unique structure. In this work, we propose an attention-guided visualization method applied to ViT that provides a high-level semantic explanation for its decision. Our method selectively aggregates the gradients directly propagated from the classification output to each self-attention, collecting the contribution of image features extracted from each location of the input image. These gradients are additionally guided by the normalized self-attention scores, which are the pairwise patch correlation scores. They are used to supplement the gradients on the patch-level context information efficiently detected by the self-attention mechanism. This approach of our method provides elaborate high-level semantic explanations with great localization performance only with the class labels. As a result, our method outperforms the previous leading explainability methods of ViT in the weakly-supervised localization task and presents great capability in capturing the full instances of the target class object. Meanwhile, our method provides a visualization that faithfully explains the model, which is demonstrated in the perturbation comparison test.