A Hybrid CNN-MLP-DWD Framework for Robust Medical Image Classification under High-Dimensional Low-Sample Size Conditions

Medical image classification in clinical settings is frequently constrained by High-Dimensional, Low-Sample Size (HDLSS) conditions, where conventional Support Vector Machines (SVMs) are geometrically susceptible to the data piling phenomenon, leading to fragile decision boundaries under feature-space perturbations. This study proposes a hybrid CNN-MLP-DWD framework that integrates multi-architecture CNN feature extraction with Distance-Weighted Discrimination (DWD) to address this statistical instability. Pre-trained ResNet50 and DenseNet121 backbones extract complementary representations, fused into a 3072-dimensional vector via Global Average Pooling and horizontal concatenation. A supervised Multi-Layer Perceptron (MLP) bottleneck then compresses this space into a 32-dimensional latent representation, resolving the computational bottleneck of deploying DWD directly on high-dimensional features. Evaluated across breast ultrasound, breast mammography, and chest X-ray datasets, the proposed framework achieves a 29-fold reduction in training latency over the baseline CNN-DWD, elevates breast ultrasound accuracy from 71.83\% to 83.93\% under HDLSS conditions, and attains a macro-AUC of 99.69\% on the chest X-ray benchmark, surpassing all compared methods. Gaussian noise perturbation tests further confirm that DWD maintains better structural resilience over SVM under out-of-distribution clinical conditions.