ROISegNet: A Region-Focused Deep Learning Framework for Precision-Driven Medical Image Segmentation

Abstract

Breast cancer is still one of the leading causes of death in women globally. Breast thermography is a non-invasive imaging procedure that can be used as an early screening device to identify temperature irregularities that suggest breast malignancies can develop. Early diagnosis is key to achieving better treatment and survival outcomes. Automated methods for breast cancer detection have drawn much attention due to the proliferation of artificial intelligence (AI) and deep learning (DL). This paper presents a new deep learning model, ROISegNet, that enables accurate segmentation of the region of interest (ROI) from breast thermography images. Our method utilizes semantic segmentation approaches using an atrous convolution-based deep learning architecture that effectively extracts the area of interest (ROI) where the abnormality is present in the human body, which potentially helps in early diagnosis. ROISegNet employs an advanced decoder module in addition to Intelligent Segmentation of Breast ROI (ISBROI), thereby achieving high segmentation accuracy. Evaluation on a benchmark dataset called DMR-IR shows that compared to other state-of-the-art backbones networks like Atrous Convolution, VGG19, ResNet50, and InceptionV3, the model's peak accuracy is 98 63%.