Scholarlink

<p>Diagnosing cervical spine fracture from CT images plays an important role in preventing long-term neurological complications. However precision and timely detection of fractures is challenging due to its complex distribution across vertebrae C1–C7. Moreover, their anatomical complexity brings additional complexities. This necessitates robust automated systems capable of detecting and differentiating the fine-grained structural abnormalities. Though traditional Deep Learning (DL) models are effective in medical image diagnosis it suffers from excessive computational loads. Also the limited interpretability of DL models makes their integration into clinical workflows particularly challenging. Moreover, existing models lack of mechanisms to select the significant features which are essential in medical image analysis. This leads to reduced generalization and diagnostic confidence. To overcome these challenges, this research work proposed MC-SELightNet with an objective of to design a lightweight, high-performance classification model that ensures accuracy and clinical transparency. The proposed MC-SELightNet integrates multi-context feature extraction through dilated depth wise convolutions with channel-wise attention enabled by Squeeze-and-Excitation (SE) blocks. This enables the network to highlight the fracture-relevant regions without increasing parameter complexity. The proposed model is specifically structured to meet clinical deployment requirements with minimized computational overhead. Also, the proposed work preserves interpretability through Grad-CAM and SE-based attention visualization. Experimental validation on the Cervical Spine Fracture Detection dataset demonstrated that the proposed model achieved higher accuracy of 98.6%, precision of 98.8%, recall of 98.2%, F1-score of 98.5%, and specificity of 99.3%, outperforming existing DL models such as DenseNet121, ResNet50, and CNN, whose maximum reported accuracy was 97.8%. "These results confirm that the proposed MC-SELightNet enhances detection precision and also delivers meaningful insights suitable for clinical interpretation and adoption.</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

A Multi-Context Squeeze-Excitation Framework with Explainable Attention for Cervical Spine Fracture Detection in CT Imaging

  • 期刊论文

  • 发表日期：

    2025年11月12日

查看全文
  • M. Anitha,
  • P. Tamije Selvy

摘要

Diagnosing cervical spine fracture from CT images plays an important role in preventing long-term neurological complications. However precision and timely detection of fractures is challenging due to its complex distribution across vertebrae C1–C7. Moreover, their anatomical complexity brings additional complexities. This necessitates robust automated systems capable of detecting and differentiating the fine-grained structural abnormalities. Though traditional Deep Learning (DL) models are effective in medical image diagnosis it suffers from excessive computational loads. Also the limited interpretability of DL models makes their integration into clinical workflows particularly challenging. Moreover, existing models lack of mechanisms to select the significant features which are essential in medical image analysis. This leads to reduced generalization and diagnostic confidence. To overcome these challenges, this research work proposed MC-SELightNet with an objective of to design a lightweight, high-performance classification model that ensures accuracy and clinical transparency. The proposed MC-SELightNet integrates multi-context feature extraction through dilated depth wise convolutions with channel-wise attention enabled by Squeeze-and-Excitation (SE) blocks. This enables the network to highlight the fracture-relevant regions without increasing parameter complexity. The proposed model is specifically structured to meet clinical deployment requirements with minimized computational overhead. Also, the proposed work preserves interpretability through Grad-CAM and SE-based attention visualization. Experimental validation on the Cervical Spine Fracture Detection dataset demonstrated that the proposed model achieved higher accuracy of 98.6%, precision of 98.8%, recall of 98.2%, F1-score of 98.5%, and specificity of 99.3%, outperforming existing DL models such as DenseNet121, ResNet50, and CNN, whose maximum reported accuracy was 97.8%. "These results confirm that the proposed MC-SELightNet enhances detection precision and also delivers meaningful insights suitable for clinical interpretation and adoption.