An Improved Network Segmentation Performance in Lesion Segmentation based on Mask R-CNN

Abstract

Computational intelligence has grown increasingly common in the analysis of healthcare data in recent years. The deep learning algorithms, in instance, have such a number of fruitful implications in the analysis of medical images. This research study has presented a double awareness deep convolution network-based system for end-to-end healthcare lesion classification. Several fully supervised and poorly supervised extracted features are included in this framework. The issue of the high cost of pixel-level labels with lesions in the medical pictures is alleviated by the weakly supervised segmentation module's use of bounding-box labels of lesion regions, which allows for correct lesion segmentation to be accomplished. In addition to this, a dual attention mechanism has been included for the purpose of improving the network's capacity for visual feature learning. The dual attention mechanism, which consists of both channel and spatial attention, may assist in drawing the network's attention to the extraction of features from significant locations. It is possible for it to considerably minimize the gaps between ground-truth labels and pseudo labels, in comparison to the current standard approach of weakly supervised segmentation using pseudo labels. Final results show that our proposed approach beat the competitors on to an oral disease collection, while our approach was extended to incorporate dermatological disease categorization in the ultimate analysis.