Clinical Pixel Feature Recalibration Module for Ophthalmic Image Classification

Ophthalmic image examination has become a commonly-acknowledged way for ocular disease screening and diagnosis. Clinical features extracted from ophthalmic images play different roles in affecting clinicians making diagnosis results, but how to incorporate these clinical features into convolutional neural network (CNN) representations has been less studied. In this paper, we propose a simple yet practical module, Clinical Pixel Feature Recalibration Module (CPF), aiming to exploit the potential of clinical features to improve the ocular disease recognition performance of CNNs. CPF first extracts clinical pixel features from each spatial position of all feature maps by clinical cross-channel pooling, then estimates each spatial position recalibration weight in a pixel-independent clinical fusion. By infusing the relative importance of clinical features into feature maps at the pixel level, CPF is supposed to enhance the representational ability of CNNs. Our CPF is easily inserted into existing CNNs with negligible overhead. We conduct comprehensive experiments on two publicly available ophthalmic image datasets and CIFAR datasets, and the results show the superiority and generation ability of CPF over advanced attention methods. Furthermore, this paper presents an in-depth weight visualization analysis to investigate the inherent behavior of CPF, aiming to improve the interpretability of CNNs in the decision-making process.