Deep neural networks have been successfully adopted to diverse domains including pathology classification based on medical images. However, large-scale and high-quality data to train powerful neural networks are rare in the medical domain as the labeling must be done by qualified experts. Researchers recently tackled this problem with some success by taking advantage of models pre-trained on large-scale general domain data. Specifically, researchers took contrastive image-text encoders (e.g., CLIP) and fine-tuned it with chest X-ray images and paired reports to perform zero-shot pathology classification, thus completely removing the need for pathology-annotated images to train a classification model. Existing studies, however, fine-tuned the pre-trained model with the same contrastive learning objective, and failed to exploit the multi-labeled nature of medical image-report pairs. In this paper, we propose a new fine-tuning strategy based on sentence sampling and positive-pair loss relaxation for improving the downstream zero-shot pathology classification performance, which can be applied to any pre-trained contrastive image-text encoders. Our method consistently showed dramatically improved zero-shot pathology classification performance on four different chest X-ray datasets and 3 different pre-trained models (5.77% average AUROC increase). In particular, fine-tuning CLIP with our method showed much comparable or marginally outperformed to board-certified radiologists (0.619 vs 0.625 in F1 score and 0.530 vs 0.544 in MCC) in zero-shot classification of five prominent diseases from the CheXpert dataset.
Weakly supervised semantic segmentation (WSSS) with only image-level supervision is a challenging task. Most existing methods exploit Class Activation Maps (CAM) to generate pixel-level pseudo labels for supervised training. However, due to the local receptive field of Convolution Neural Networks (CNN), CAM applied to CNNs often suffers from partial activation -- highlighting the most discriminative part instead of the entire object area. In order to capture both local features and global representations, the Conformer has been proposed to combine a visual transformer branch with a CNN branch. In this paper, we propose TransCAM, a Conformer-based solution to WSSS that explicitly leverages the attention weights from the transformer branch of the Conformer to refine the CAM generated from the CNN branch. TransCAM is motivated by our observation that attention weights from shallow transformer blocks are able to capture low-level spatial feature similarities while attention weights from deep transformer blocks capture high-level semantic context. Despite its simplicity, TransCAM achieves a new state-of-the-art performance of 69.3% and 69.6% on the respective PASCAL VOC 2012 validation and test sets, showing the effectiveness of transformer attention-based refinement of CAM for WSSS.
Contrastive learning has led to substantial improvements in the quality of learned embedding representations for tasks such as image classification. However, a key drawback of existing contrastive augmentation methods is that they may lead to the modification of the image content which can yield undesired alterations of its semantics. This can affect the performance of the model on downstream tasks. Hence, in this paper, we ask whether we can augment image data in contrastive learning such that the task-relevant semantic content of an image is preserved. For this purpose, we propose to leverage saliency-based explanation methods to create content-preserving masked augmentations for contrastive learning. Our novel explanation-driven supervised contrastive learning (ExCon) methodology critically serves the dual goals of encouraging nearby image embeddings to have similar content and explanation. To quantify the impact of ExCon, we conduct experiments on the CIFAR-100 and the Tiny ImageNet datasets. We demonstrate that ExCon outperforms vanilla supervised contrastive learning in terms of classification, explanation quality, adversarial robustness as well as calibration of probabilistic predictions of the model in the context of distributional shift.
Despite the success of deep learning in computer vision, algorithms to recognize subtle and small objects (or regions) is still challenging. For example, recognizing a baseball or a frisbee on a ground scene or a bone fracture in an X-ray image can easily result in overfitting, unless a huge amount of training data is available. To mitigate this problem, we need a way to force a model should identify subtle regions in limited training data. In this paper, we propose a simple but efficient supervised augmentation method called Cut\&Remain. It achieved better performance on various medical image domain (internally sourced- and public dataset) and a natural image domain (MS-COCO$_s$) than other supervised augmentation and the explicit guidance methods. In addition, using the class activation map, we identified that the Cut\&Remain methods drive a model to focus on relevant subtle and small regions efficiently. We also show that the performance monotonically increased along the Cut\&Remain ratio, indicating that a model can be improved even though only limited amount of Cut\&Remain is applied for, so that it allows low supervising (annotation) cost for improvement.
Recent years have seen the introduction of a range of methods for post-hoc explainability of image classifier predictions. However, these post-hoc explanations may not always align perfectly with classifier predictions, which poses a significant challenge when attempting to debug models based on such explanations. To this end, we seek a methodology that can improve alignment between model predictions and explanation method that is both agnostic to the model and explanation classes and which does not require ground truth explanations. We achieve this through a novel explanation-driven data augmentation (EDDA) method that augments the training data with occlusions of existing data stemming from model-explanations; this is based on the simple motivating principle that occluding salient regions for the model prediction should decrease the model confidence in the prediction, while occluding non-salient regions should not change the prediction -- if the model and explainer are aligned. To verify that this augmentation method improves model and explainer alignment, we evaluate the methodology on a variety of datasets, image classification models, and explanation methods. We verify in all cases that our explanation-driven data augmentation method improves alignment of the model and explanation in comparison to no data augmentation and non-explanation driven data augmentation methods. In conclusion, this approach provides a novel model- and explainer-agnostic methodology for improving alignment between model predictions and explanations, which we see as a critical step forward for practical deployment and debugging of image classification models.
Visualizing the features captured by Convolutional Neural Networks (CNNs) is one of the conventional approaches to interpret the predictions made by these models in numerous image recognition applications. Grad-CAM is a popular solution that provides such a visualization by combining the activation maps obtained from the model. However, the average gradient-based terms deployed in this method underestimates the contribution of the representations discovered by the model to its predictions. Addressing this problem, we introduce a solution to tackle this issue by computing the path integral of the gradient-based terms in Grad-CAM. We conduct a thorough analysis to demonstrate the improvement achieved by our method in measuring the importance of the extracted representations for the CNN's predictions, which yields to our method's administration in object localization and model interpretation.