ROISegNet: Anatomy-Aware Deep Learning Framework for Accurate Breast ROI Segmentation in Thermographic Imaging

Abstract

Breast thermography (BT) is a new imaging modality that is a safe, non-invasive method with potential usefulness of early breast abnormalities detection. But the clinical value is restricted by the challenge of reproducibly determining anatomic relevant region of interests (ROIs) under differing thermal conditions, as well as background distractions. This problem indicates the space of automated and strong ROI segmentation techniques. In this work, we propose a sematic segmentation framework based on deep learning specifically for breast thermogram analysis called ROISegNet. Our architecture consists of a hierarchical atrous convolution-based encoder for multi-scale feature extraction, improved atrous spatial pyramid pooling (ASPP) module for context information capturing and a novel anatomy guided ROI localisation strategy ISBROI to reduce irrelevant background pixels and localise breast boundaries. Using the same experimental protocol, the model is evaluated on the publicly accessible DMR-IR dataset and compared with previously established architectures such as VGG19, ResNet50, InceptionV3, and other atrous convolution-based approaches. Experimental results showed that ROISegNet gives promising, consistent, and stable performances with the average accuracy of∼97%, average mean IoU (93–94%), precision, and less Dice loss. These results suggest bounds preservation efficacy and alignment in space. Lastly, statistical validation across numerous experimental runs validates the stability and reproducibility of the proposed framework. We conclude that ROISegNet allows for semi-automatic accurately, anatomically consistently segmentation of breast ROIs, making it a promising preprocessing component in computer-aided analysis systems for thermography. This approach could allow more reliable non-invasive breast screening once further optimisation and validation of the methodology is complete, thereby facilitating early clinical diagnosis.