Optimized AlexNet Pruning for Edge-Based Medical Diagnostics

Medical diagnostics demand rapid and accurate disease detection to ensure timely treatment, directly impacting human lives. Deep neural networks (DNNs) have shown unparalleled success in medical applications, surpassing traditional methods. However, their computational and memory requirements often exceed the capabilities of resource-constrained edge devices, such as mobile and hand held platforms, limiting their deployment in real-world scenarios. To address this, model compression techniques have emerged as a key solution. This paper focuses on pruning AlexNet to significantly reduce its size. The reduced network effectively processes the PAD-UFES-20 dataset, distinguishing melanoma, a serious type of skin cancer, from benign skin lesions, while maintaining high classification accuracy. A meticulous layer-by-layer investigation evaluates magnitude-based pruning at varying ratios, analyzing its effects on accuracy and resource efficiency. The results reveal a clear difference between fully connected (FC) and convolutional layers: pruning FC layers substantially reduces memory consumption, while pruning convolutional layers significantly boosts inference speed. Careful analysis resulted in reaching 91% pruning, which resulted in accuracy dropping from 99.16% to 97.48%, with big reduction in memory usage. This balance demonstrates the feasibility of deploying pruned DNNs for real-time, edge-based medical diagnostics without compromising diagnostic precision. Furthermore, the proposed pruning approach was also applied to the full PAD-UFES-20 dataset across all classes, yielding similar results to the two-class (melanoma vs. benign) case. This work provides a robust framework for optimizing DNNs for resource-constrained environments, offering valuable insights for designing efficient, practical models in medical applications.