A ranking-based multi-scale feature calibration network for nuclear cataract grading in AS-OCT images

Anterior Segment Optical Coherence Tomography (AS-OCT) is an important imaging technique for the grading of nuclear cataract. However, due to the complex interdependencies among 6 clinically-defined levels of cataract severity, it presents significant challenges to classify neighboring severity levels accurately and expeditiously, whether by human experts or computer-aided approaches. Existing deep learning-based models usually obtain 3 grades of nuclear cataract severity only, and often struggle to capture vital information related to the progression of neighboring severity levels, leading to inaccuracies in grading. In this paper, we introduce a novel method called Ranking-MFCNet, which utilizes both a ranking-based framework and a Multi-scale Feature Calibration network (MFCNet). To bolster the model's capability for discriminating between neighboring severities that are prone to confusion, we treat the multi-category severity classification as a collection of distinct binary classification patterns. This strategy facilitates a systematic implementation of fine-grained nuclear cataract severity grading on an individual basis. Within each binary classification pattern, we propose an external attention-augmented Multi-scale Feature Calibration (eaMFC) module, which effectively captures the multi-scale characteristics inherent to the lens nucleus. Additionally, eaMFC allows for the calibration of shared attributes extracted by the external attention layer, thereby enhancing the model's proficiency in modeling the distinctive traits related to opacity and sclerosis of the lens nucleus. We trained and validated our model on a dataset that contains 1608 AS-OCT images, and the extensive experiments have verified the effectiveness and superiority of our method over state-of-the-art cataract grading methods.