Advancing Medical Image Classification: Harnessing Additive Angular Margin Loss for Enhanced Discrimination

The shortage of medical specialists across the globe poses a significant challenge, and there is a need for innovative solutions to address this gap. Biomedical imagery is a precious asset in the healthcare and life sciences, with machine learning techniques significantly propelling improvements at the forefront of image classification. This research work investigates the scope of the additive angular margin loss to enhance discriminative capability among inter-classes and improve classification accuracy on medical images. We trained the ResNet-18 network with a set of hyperparameters of the loss function to visualize the impact of the interplay of margin (m) and scale (s) on the MedMNIST (v2) dataset. Our experiments revealed that specific values for 'm'and's' could significantly improve feature discrimination in particular image datasets that depend on the total target classes, mainly when applied judiciously. Our research suggests that the additive angular margin loss function is an assuring method for medical image classification. For medical image analysis, where tiny variations can be crucial, creating clear and distinct representations is vital to accurately identifying different diseases or categories. Overall, this research contributes to the growing body of knowledge on machine learning for medical image classification, and its findings pave the way for future studies to develop more robust and generalizable algorithms.