An ensemble attention-based feature guidance UNet++ with customized loss for brain tumor MRI segmentation

A brain tumor is one of a clinically significant diseases as it can lead to serious clinical implications can adversely affect survival without timely treatment. Therefore, the accurate and timely diagnosis of brain tumor regions is a major challenge in computational medicine. With the introduction of deep learning in the healthcare sector, improvements have been made in this regard; however, these models still have issues with sensitive areas and boundaries in tumor regions. This problem arises primarily from class imbalance, which leads to reduced performance. In this study, we present a new approach for optimizing medical image segmentation based on the UNet++ architecture. The proposed algorithm focuses on identifying the most informative image regions, which results in more precise segmentation and reliable assessment. We introduce a hybrid loss function, a composition of Binary Cross-Entropy and Focal Tversky loss. This composition helps to improve segmentation accuracy and the class imbalance problem. The Binary Cross-Entropy component ensures consistent and accurate pixel-wise classification, while the Focal Tversky term strengthens the delineation of tumors under severe class imbalance. Collectively, these features allow the model to learn in a way that is not sensitive to the challenges in areas of diagnostics and does not reduce the overall segmentation integrity. In order to further increase the strength and flexibility of the model, we use a hybrid validation approach combining K-fold cross-validation with test-time augmentation, augmented with a new weighted ensemble mechanism. The ensemble combines the predictions of the most competent models, which results in the final segmentation results, which are more consistent and accurate. In general, the experimental framework shows that the complementary benefits of incorporating specific architectural improvements, the design of loss function, and efficient ensemble learning are central to a stable and effective brain-tumor segmentation of medical images.