Robust medical image classification with curvature regularization on the PATHMNIST

In medical image analysis, the robustness and accuracy of classification models are paramount, especially under adverse conditions such as data scarcity, class imbalance, and potential adversarial attacks. This paper introduces a novel deep learning architecture employing curvature regularization (CURE) to enhance the robustness of medical image classification on the PATHMNIST dataset. Our approach integrates advanced convolutional techniques including depthwise and dilated convolutions with an attention mechanism, focusing on precise and detailed feature extraction essential for medical diagnostics. We incorporate curvature regularization to stabilize the learning process by controlling the magnitude of the Hessian's eigenvalues, making the model less sensitive to input variations. The effectiveness of our architecture is demonstrated through rigorous testing, including adversarial scenarios using the Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD). Results show our model not only achieves superior accuracy compared to traditional architectures like ResNet-18 and ResNet-50 but also maintains higher resilience against adversarial attacks. This robustness is critical for practical deployment in medical image analysis.