Enhancing medical image segmentation with parameter-efficient involutional neural networks and diverse datasets

Abstract

Medical image segmentation is crucial for precise analysis and efficient treatment schemes. The utilization of Involutional Neural Networks for the enhancement of executing segmentation by lessening parameter complexity and enhancing feature extraction adaptability is initiated by this research. Traditional convolutional techniques are surpassed by Involutional Neural Networks as it offers dynamic adjustments which makes involution more suitable for tasks where computational efficiency is critical. The potential of Involutional Neural Networks and diversified datasets in improving medical image processing and introducing more definitive and efficient medical aid systems is highlighted in this research. Grad-CAM visualization has been integrated as a part of this model which generated heatmaps to focus on the critical regions, clarifying its decision-making process. In this research four datasets were used, which were the BUSI dataset for breast cancer segmentation, the DRIVE dataset for retinal vessel segmentation, the Lung Image Segmentation dataset for segmenting the human lungs, and the HAM-10000 dataset for skin lesion segmentation. Out of these four datasets, the model performed the best with the DRIVE dataset scoring an IoU of 97.07% and an accuracy of 96.76%. The HAM-10000 dataset scored an IoU of 77.71% with an accuracy of 97.85%, the BUSI dataset scored an IoU of 70.23% with an accuracy of 93.5% and finally, the Lung Image Segmentation dataset scored an IoU of 78.68% with an accuracy of 98.01%.