Analysis of Breast Region Segmentation in Thermal Images us-ing U-Net Deep Neural Network Variants

Mohamed Hadi Habaebi^1*Rafhanah Shazwani¹Mohammed Al Hussaini²Md Rafiqul Islam¹

• ¹International Islamic University Malaysia, Selayang, Malaysia
• ²Arab Open University, Muscat, Oman

Breast cancer detection using thermal imaging requires reliable region segmentation techniques for the precise delineation of the breast region from proximal body areas. In this study, three different segmentation models, namely U-Net, U-Net with Spatial Attention, and U-Net++, were comparatively analyzed using five optimization algorithms, which are ADAM, NADAM, RMSPROP, SGDM, and ADADELTA. Evaluation, conducted through k-fold cross-validation, includes metrics such as Intersection over Union, Dice coefficient, precision, recall, sensitivity, specificity, pixel accuracy, ROC-AUC, PR-AUC, and integrated Grad-CAM heatmaps for qualitative analysis. The results revealed that the ADAM optimizer outperforms the other optimizers, producing superior accuracy and reduced loss in this study. Despite the emergence of complex models, foundational U-Net, trained with the ADAM optimizer, continues to demonstrate effectiveness. The findings indicate that there is not always a direct correlation between architectural complexity and improved out-comes. U-Net demonstrated impressive precision (0.9721), recall (0.9559), and specificity (0.9801). Its discriminative ability was reflected in ROC-AUC (0.9680) and PR-AUC (0.9472). U-Net also demonstrates superior performance in breast region overlap and noise handling. This research con-tributes valuable insights for medical image analysis, guiding the selection of appropriate models and optimizers for similar applications, ultimately enhancing the efficiency and accuracy of breast cancer diagnosis using thermal imaging.