A carregar...
Publicação
Automatic detection of lumbar spinal stenosis in computed magnetic resonance imaging
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 6.88 MB | Adobe PDF |
Autores
Orientador(es)
Resumo(s)
LSS is a common degenerative spinal condition characterized by narrowing of the spinal canal, which can result in pain, functional impairment, and a diminished quality of life. Despite the availability of advanced imaging techniques, such as Magnetic Resonance Imaging (MRI), the diagnosis and severity classification of LSS remain challenging due to subjective interpretation, inter-observer variability, and the absence of standardized assessment criteria. These challenges underscore the need for automated and objective
tools to support clinical decision-making. This thesis proposes a computer vision and deep learning framework for the automated
detection and classification of LSS in sagittal lumbar MRI scans. The framework adopts a two-stage approach: first, vertebrae are localized using a YOLOv8 object detection model, and second, regions of interest at the intervertebral disc level are extracted and
classified by a Swin Transformer-based model into three severity categories: Normal/Mild, Moderate, and Severe. Data preprocessing steps included filtering, resizing, contrast enhancement, and manual annotation of vertebrae to ensure high-quality inputs. The
framework was developed and validated using the RSNA 2024 multi-institutional dataset, which contains diverse lumbar spine MRI studies with standardized severity labels. Experimental evaluation demonstrated that the YOLOv8 detection model achieved high performance, with 97.83% precision, 98.03% recall, and an F1-score of 97.93% in vertebra localization. The Swin Transformer-based classifier achieved a weighted F1- score of 77.05% and a macro F1-score of 67.98%, while the integrated framework yielded a
weighted F1-score of 77.15%. Additionally, the system demonstrated clinical feasibility by processing each study in an average of 1.64 seconds, supporting its potential for integration into diagnostic workflows. The findings of this study confirm that Artificial Intelligence (AI)-driven methods can enhance diagnostic accuracy and efficiency in LSS assessment. Limitations, such as restricted dataset diversity and the need for prospective clinical validation, are discussed, along with future research directions, including the use of multi-plane imaging, larger annotated datasets, and advanced transformer-based architectures.
Descrição
Palavras-chave
Lumbar Spinal Stenosis (LSS) Magnetic Resonance Imaging (MRI) Deep Learning YOLOv8 Swin Transformer Convolutional Neural Networks (CNN) Object Detection Medical Image Classification Clinical Decision Support Artificial Intelligence in Healthcare