Analysis of breast region segmentation in thermal images using U-Net deep neural network variants.

Introduction: Breast cancer detection using thermal imaging relies on accurate segmentation of the breast region from adjacent body areas. Reliable segmentation is essential to improve the effectiveness of computer-aided diagnosis systems. Methods: This study evaluated three segmentation models-U-Net, U-Net with Spatial Attention, and U-Net++-using five optimization algorithms (ADAM, NADAM, RMSPROP, SGDM, and ADADELTA). Performance was assessed through k-fold cross-validation with metrics including Intersection over Union (IoU), Dice coefficient, precision, recall, sensitivity, specificity, pixel accuracy, ROC-AUC, PR-AUC, and Grad-CAM heatmaps for qualitative analysis. Results: The ADAM optimizer consistently outperformed the others, yielding superior accuracy and reduced loss. Among the models, the baseline U-Net, despite being less complex, demonstrated the most effective performance, with precision of 0.9721, recall of 0.9559, specificity of 0.9801, ROC-AUC of 0.9680, and PR-AUC of 0.9472. U-Net also achieved higher robustness in breast region overlap and noise handling compared to its more complex variants. The findings indicate that greater architectural complexity does not necessarily lead to improved outcomes. Discussion: This research highlights that the original U-Net, when trained with the ADAM optimizer, remains highly effective for breast region segmentation in thermal images. The insights contribute to guiding the selection of suitable deep learning models and optimizers for medical image analysis, with the potential to enhance the efficiency and accuracy of breast cancer diagnosis using thermal imaging.