Attention-based multiple instance learning for lung neoplasm detection using scleral images

Abstract

Lung cancer is one of the most fatal malignant tumors globally. Recent research has revealed that the scleral image, which can be obtained in a painless and non-invasive manner conveniently, is associated with lung cancer. The current method treats the malignant lung neoplasm detection task as a binary classification problem. It takes the multiple images of one subject's eyes as different instances, and takes the subject as a bag, then employs the multiple instance learning technique to solve the problem. However, the current method utilizes average pooling to aggregate information on different instances in a bag, which overlooks the varying contribution of each instance to the bag label. In this study, we propose an attentionbased multiple instance learning (AMIL) model for lung neoplasm detection using scleral images. The model first employs a convolutional neural network-based backbone to extract features from each scleral image of the subject, then leverages a channel attention module to recalibrate the weight and aggregate the extracted features adaptively. Finally, a fully connection-based classification module is used to make the final prediction for the subject. The results show that the model can effectively identify the critical instances with the help of the attention mechanism, and improve the classification accuracy.