Regresar

An Architecture for MRI Processing, Segmentation and Model Explanation Using Deep Learning and Transfer Learning to Support Brain Cancer Diagnosis in Preoperative Patients

Abstract:

Brain cancer diagnosis, particularly in preoperative stages, presents a complex challenge due to the heterogeneity of brain lesions. This work introduces a novel and comprehensive architecture that combines advanced deep learning techniques for MRI image preprocessing, segmentation, and model explanation, tailored specifically for preoperative brain tumor analysis. Unlike previous studies that focus on postoperative data or missing modality synthesis, our approach integrates multiple MRI modalities (T1c, T2w, FLAIR) and incorporates advanced features such as residual connections, attention mechanisms, and multi-scale inputs in 3D U-Net architectures. A key innovation of our work is the introduction of model explainability, which enhances clinical interpretability and trust in the model’s predictions—an aspect often overlooked in traditional segmentation approaches. Our architecture follows a five-phase process: data normalization and dimensionality reduction, an optimized image processing pipeline, the training of multiple 3D U-Net variants, fine-tuning based on performance metrics such as Dice coefficient and Hausdorff distance, and, crucially, the explainability of the models. The model with three MRI modalities consistently outperformed others, demonstrating superior precision and robustness. By addressing both the accuracy of brain tumor segmentation and the need for explainability in clinical settings, this work offers a significant advancement over traditional methods. Future work will focus on refining edge delineation and further integrating these models into clinical workflows, enhancing both performance and trustworthiness in real-world applications.