On the surprising accuracy, efficiency, and interpretability of shallow neural networks in medical image classification

Over the past decade, a major inspiration for research and application in medicine has been the rapid development of artificial intelligence (AI) tools, particularly in computer vision, resulting in the common belief that deeper neural networks tend to provide better performance. The concepts and strategies developed in AI have often been applied directly to medical data without considering their unique characteristics and requirements. Despite promising results, particularly in imaging and diagnosis, several challenges remain that prevent widespread clinical adoption. These include high computational requirements and limited model interpretability of complex modes, as well as issues associated with data scarcity. The comparative analysis presented in this thesis examines shallow models’ performance against deep networks on three different datasets, PneumoniaMNIST, BreastMNIST and ADNI. A deep model, ResNet-50, was tested against shallow networks designed in various architectural configurations for this study, and the integration of texture features extracted via PyRadiomics and edge features derived from image derivation maps was investigated to assess their influence on model accuracy and interpretability. Contrary to common assumptions, the results indicate that shallow models can achieve comparable or even superior performance, improving accuracy by up to 27% while significantly reducing computational cost and improving interpretability. In addition, the study highlights the need for standardization and optimized resource allocation to support broader research efforts, thereby increasing the potential for the development of impactful tools to improve patient care.