A COMPARATIVE STUDY OF DEEP LEARNING-BASED SEGMENTATION TECHNIQUES: U-NET, SEGNET, AND DEEPLABV3 FOR BREAST CANCER DETECTION BY THERMAL IMAGING

Image segmentation is vital for a wide range of computer vision applications, where precise delineation of regions within an image is essential. The breast thermal images acquired are segmented manually using various tools such Photoshop, label box. Images can also be segmented semi automatically using point-based annotation, a bounding box, etc. As the, semi-automatic methods are time-consuming with large datasets, a fully automatic segmentation model is required to provide a better solution. To address this, pretrained models with various backbone networks can be used with a few manually annotated masks to segment the images. The three different architecture models, namely U-net, SegNet, and a pretrained DeepLabv3 were analyzed, using the datasets holding 1000 images and their corresponding mask (ground truth). The three models have been compared based on accuracy, pixel accuracy, dice score and Intersection over Union. U-Net yielded better results, achieving a dice coefficient of 96 %, and it required less training time than the SegNet, which managed a dice score of 93%. DeepLabv3 showed relatively lower performance with a Dice Coefficient of 89%.