Segmentação de múltiplas patologias em ressonância magnética da coluna lombar: uma abordagem comparativa de aprendizado profundo

Abstract

Low back pain is a leading cause of disability worldwide, and Magnetic Resonance Imaging (MRI) of the lumbar spine is fundamental for its diagnosis. However, the manual analysis of these images is a time-consuming and subjective process, and existing computational methods are often limited to a single pathology or rely on complex, multi-stage pipelines. A central challenge, not yet fully addressed, is the simultaneous occurrence of multiple pathologies in the same anatomical structure—a common clinical scenario that current models fail to model effectively. This dissertation addresses this gap by proposing and validating a robust methodology for the automated segmentation of multiple co-existing pathologies in lumbar intervertebral discs. The work was structured in two complementary phases. The first established the methodological underpinnings of this research. It consisted of a rigorous empirical study evaluating five deep learning architectures and four loss functions to determine the most effective approaches for the fundamental task of segmenting the vertebrae and intervertebral discs. This step ensured that the subsequent investigation into pathologies was built upon a robust and validated base. The second phase, the main contribution of this work, systematically investigated three distinct strategies for multi-pathology segmentation: (i) binary class segmentation, a baseline that treats each pathology independently; (ii) multi-class segmentation, mapping 70 disease combinations to unique classes (non-overlapping masks); and (iii) multi-label segmentation, which uses binary channels to explicitly model the coexistence of multiple diagnoses (overlapping masks). Our results, derived from over 200 training pipelines, demonstrate that the multi-label approach, especially when implemented with the V-Net and Swin UNETR architectures, achieves diagnostic accuracy comparable to the baseline while offering significantly superior computational efficiency. By developing a unified framework that integrates the precise spatial localization of symptomatic areas with the classification of multiple concurrent diseases, this work establishes a practical and efficient guideline for future research and clinical applications in the automated diagnosis of spinal pathologies.