Scholarlink

Purpose <p>The study aims to develop a deep learning-based method for the automatic identification of pathological vertebrae in lumbar spine MRI scans.</p> Methods <p>A retrospective cohort of 98 subjects was used, including 31 healthy controls and 67 patients with vertebral pathologies (infection, inflammation, or carcinoma). T2-weighted MRI scans were acquired from 1.5T and 3.0T MRI scanners. A RetinaNet object detection algorithm with a ResNet18 backbone was used for vertebrae localization in 2D T2-weighted MRI images. The detected vertebrae were classified as healthy or pathological using a ConvNeXt model, with hyperparameter optimization performed using Optuna.</p> Results <p>The object detection model achieved an Intersection-Over-Union (IOU) score of 0.67 for pathological vertebrae. The 2D-based classification model, enhanced by post-processing using a LightGBM tree-based model, achieved a test accuracy of 95.38%, with precision of 86.67%, recall of 92.86%, and an F1-score of 89.66%.</p> Conclusions <p>This study demonstrates the feasibility of deep learning models for detecting pathological vertebrae in lumbar spine MRI scans. The integration of 2D and 3D imaging data enhances diagnostic accuracy and presents an automated radiologist assistant tool for improved spinal pathology diagnosis, potentially streamlining clinical workflows.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Automatic detection of spinal pathologies based on MRI scans

  • 期刊论文

  • 发表日期：

    2025年11月19日

查看全文
  • Omer Dor,
  • Oz Haim,
  • Lee Azolai,
  • Ariel Agur,
  • May Cohen,
  • Noam Shomron,
  • Moran Artzi

摘要

Purpose

The study aims to develop a deep learning-based method for the automatic identification of pathological vertebrae in lumbar spine MRI scans.

Methods

A retrospective cohort of 98 subjects was used, including 31 healthy controls and 67 patients with vertebral pathologies (infection, inflammation, or carcinoma). T2-weighted MRI scans were acquired from 1.5T and 3.0T MRI scanners. A RetinaNet object detection algorithm with a ResNet18 backbone was used for vertebrae localization in 2D T2-weighted MRI images. The detected vertebrae were classified as healthy or pathological using a ConvNeXt model, with hyperparameter optimization performed using Optuna.

Results

The object detection model achieved an Intersection-Over-Union (IOU) score of 0.67 for pathological vertebrae. The 2D-based classification model, enhanced by post-processing using a LightGBM tree-based model, achieved a test accuracy of 95.38%, with precision of 86.67%, recall of 92.86%, and an F1-score of 89.66%.

Conclusions

This study demonstrates the feasibility of deep learning models for detecting pathological vertebrae in lumbar spine MRI scans. The integration of 2D and 3D imaging data enhances diagnostic accuracy and presents an automated radiologist assistant tool for improved spinal pathology diagnosis, potentially streamlining clinical workflows.