Scholarlink

Purpose <p>Meniscus injuries are common intra-articular knee pathologies caused by trauma, degeneration, or overuse, often leading to pain, swelling, and restricted mobility. Early and accurate diagnosis of meniscal injuries is essential to prevent long-term joint damage. Despite MRI being the gold standard, manual interpretation is time-consuming and prone to variability, particularly when differentiating subtle tears from degeneration. This study proposes a radiologist-inspired framework integrating preprocessing and lightweight hybrid classification using MobilenetV3 as feature extractor and Radial basis function – Support vector machine (RBF- SVM) as classifier to improve diagnostic performance.</p> Methods <p>Sagittal fat-suppressed knee MRI images were pre-processed using an <i>empirically derived formula</i> to enhance visibility of linear hyperintensities (tears), diffuse patches (degeneration), and their co-occurrence. A two-stage <i>hierarchical classification</i> pipeline was implemented: binary classification (Normal vs. Diseased) followed by multi-class classification (Tear, Degeneration, Tear with Degeneration) using MobileNetV3 for feature extraction and an RBF-SVM for classification. Grad-CAM was applied for interpretability analysis.</p> Results <p>The proposed framework achieved an AUC of 1.0 for binary and &gt; 0.98 for multi-class classification. Accuracy improved from 73.88% on raw data to 95.75% after preprocessing. Sensitivity, specificity, precision, and F1-score for multi-class classification were 96.12%, 95.40%, 95.88%, and 95.60%, respectively, demonstrating balanced performance across all categories. Grad-CAM confirmed model attention on the meniscus region, consistent with radiologist focus.</p> Conclusion <p>The proposed method demonstrates high accuracy, speed, and interpretability through efficient feature extraction and robust classification, with a lightweight and computationally efficient design enabling real-time clinical deployment. However, Grad-CAM visualizations confirm model focus on the meniscus region to support interpretability.</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

Radiologist-Inspired Meniscus Injury Detection Using MobileNetV3-SVM with Grad-CAM Visualization

  • 期刊论文

  • 发表日期：

    2025年11月17日

查看全文
  • Priya Choudhary,
  • Abha Jaiswal,
  • Debadutta Dash,
  • Ashish Verma,
  • Shiru Sharma,
  • Neeraj Sharma

摘要

Purpose

Meniscus injuries are common intra-articular knee pathologies caused by trauma, degeneration, or overuse, often leading to pain, swelling, and restricted mobility. Early and accurate diagnosis of meniscal injuries is essential to prevent long-term joint damage. Despite MRI being the gold standard, manual interpretation is time-consuming and prone to variability, particularly when differentiating subtle tears from degeneration. This study proposes a radiologist-inspired framework integrating preprocessing and lightweight hybrid classification using MobilenetV3 as feature extractor and Radial basis function – Support vector machine (RBF- SVM) as classifier to improve diagnostic performance.

Methods

Sagittal fat-suppressed knee MRI images were pre-processed using an empirically derived formula to enhance visibility of linear hyperintensities (tears), diffuse patches (degeneration), and their co-occurrence. A two-stage hierarchical classification pipeline was implemented: binary classification (Normal vs. Diseased) followed by multi-class classification (Tear, Degeneration, Tear with Degeneration) using MobileNetV3 for feature extraction and an RBF-SVM for classification. Grad-CAM was applied for interpretability analysis.

Results

The proposed framework achieved an AUC of 1.0 for binary and > 0.98 for multi-class classification. Accuracy improved from 73.88% on raw data to 95.75% after preprocessing. Sensitivity, specificity, precision, and F1-score for multi-class classification were 96.12%, 95.40%, 95.88%, and 95.60%, respectively, demonstrating balanced performance across all categories. Grad-CAM confirmed model attention on the meniscus region, consistent with radiologist focus.

Conclusion

The proposed method demonstrates high accuracy, speed, and interpretability through efficient feature extraction and robust classification, with a lightweight and computationally efficient design enabling real-time clinical deployment. However, Grad-CAM visualizations confirm model focus on the meniscus region to support interpretability.