Enhancing Model Efficiency in Medical Image Analysis through Combined Stripe-Wise and Filter Pruning Techniques

Abstract

With the development of deep learning techniques in recent years, computerised medical diagnosis has advanced significantly. Putting a deep learning model to use for portable and inexpensive devices is a significant bottleneck, though. In order to have real-time applica tions and models that can be deployed on resource-constrained EDGE devices, the growth of the medical profession necessitates that we be able to reduce the inference time with these approaches. This research investigates the effectiveness of filter pruning methods in optimiz ing the architecture of DNNs for medical image classification, for the HAM10000 skin cancer and Diabetes Retinopathy Detection (DRD) datasets. Given the large size of neural networks like VGG-13 and ResNet-56, which are computationally intensive and difficult to deploy in resource-constrained environments, there’s a need to optimize these networks while re taining their predictive performance. The experiments we did demonstrate that despite a substantial decrease in the size of the model, the number of FLOPs and model parameters, the pruned versions of both the VGG-13 and ResNet-56 models maintained a comparable level of accuracy to the original, non-pruned versions, and in some cases, even surpassed the original accuracy (ResNet-56 on DRD dataset). These results suggest that filter pruning can be an effective method for reducing the computational and storage requirements of neu ral networks in the medical field without compromising their diagnosing performance. The implementation of pruned models could enhance the accessibility and usability of these neu ral networks in real-world clinical settings where computational resources may be limited.