Scholarlink

The growing demand for lumbar spine MRI exams, coupled with a shortage of radiologists, highlights the need for automated diagnostic tools for spinal conditions. Despite advancements in AI across radiology, no CE or FDA-approved solutions currently target lumbar spine pathologies. This paper introduces a deep learning pipeline designed to automatically detect intervertebral disc herniation, aiming to support faster and more accurate clinical decisions. A dataset of 165 lumbar spine MRI exams, totalling 5,200 sagittal slices from four manufacturers, was annotated by radiologists. The experiment includes a multi-stage approach. MA-Net with EfficientNet-B2 achieved the best segmentation results, reaching a Dice Score of 0.898. For herniation classification, the ViT model outperformed others, achieving an F1-Score of 0.905 and accuracy of 0.826. Simplifying the segmentation task to three classes enhanced robustness, and anatomical labelling enabled precise disc-level classification. The pipeline demonstrates the feasibility of automated herniation detection in lumbar MRI, supporting improved diagnostic consistency and reduced radiologist burden.

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

Deep Neural Networks to the Detection of Lumbar Hernias: Methodology and Preliminary Results

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • António Fernandes,
  • João Rodriguez,
  • Susana Moleirinho,
  • Irina Trofimenko,
  • Ekaterina Guseva,
  • Alexander Martinovich,
  • Ilzane Morais,
  • Louise Bisolo,
  • Luis M. Gomes,
  • José M. Machado

摘要

The growing demand for lumbar spine MRI exams, coupled with a shortage of radiologists, highlights the need for automated diagnostic tools for spinal conditions. Despite advancements in AI across radiology, no CE or FDA-approved solutions currently target lumbar spine pathologies. This paper introduces a deep learning pipeline designed to automatically detect intervertebral disc herniation, aiming to support faster and more accurate clinical decisions. A dataset of 165 lumbar spine MRI exams, totalling 5,200 sagittal slices from four manufacturers, was annotated by radiologists. The experiment includes a multi-stage approach. MA-Net with EfficientNet-B2 achieved the best segmentation results, reaching a Dice Score of 0.898. For herniation classification, the ViT model outperformed others, achieving an F1-Score of 0.905 and accuracy of 0.826. Simplifying the segmentation task to three classes enhanced robustness, and anatomical labelling enabled precise disc-level classification. The pipeline demonstrates the feasibility of automated herniation detection in lumbar MRI, supporting improved diagnostic consistency and reduced radiologist burden.