Yang Guang, Shi Yubo, Li Qiang
School of Microelectronics, Tianjin University, Tianjin, China.
Department of Orthopedics Surgery, Second Hospital of Tianjin Medical University, Tianjin Medical University, Tianjin, China.
Quant Imaging Med Surg. 2025 Jun 6;15(6):5811-5824. doi: 10.21037/qims-24-1808. Epub 2025 Jun 3.
Anterior cruciate ligament (ACL) and meniscal injuries are commonly diagnosed with multi-plane magnetic resonance imaging (MRI), but most artificial intelligence (AI) models use single-plane input, leaving the roles of each plane underexplored. This study aimed to investigate the differential impacts and interactions of sagittal, coronal, and axial knee MRI planes on the detection of ACL tears and meniscal tears within a deep learning (DL) model.
The MRNet dataset, consisting of 1,130 training cases and 120 validation cases, was employed to develop the TripleMRNet model. This model was trained on images from one, two, or three planes, resulting in seven combinations. This study systematically compared diagnostic performance across these combinations with gradient-weighted class activation mapping (Grad-CAM) providing interpretability analysis.
For ACL tear detection, the three-plane model demonstrated the highest performance, achieving an accuracy (ACC) of 0.925, sensitivity (SEN) of 0.944, specificity (SPE) of 0.909, and F1 score of 0.919. The coronal model had the lowest ACC (0.842), SPE (0.833), and F1 score (0.829). For meniscal tear detection, although SEN remained similar across all seven models, the 3-plane model demonstrated superior performance in terms of ACC (0.783), SPE (0.824), and F1 score (0.745). The axial model ranked just below the 3-plane model across these three metrics, with only a slight margin. Conversely, the sagittal model performed the worst, with an ACC of 0.633, SPE of 0.545, and an F1 score of 0.639.
The sagittal plane was shown to be the most effective for detecting ACL tears, with the axial MR images also demonstrating significant utility. For meniscal tear detection, the axial plane markedly outperformed the other two planes, and the sagittal plane exhibited the poorest performance.
前交叉韧带（ACL）和半月板损伤通常通过多平面磁共振成像（MRI）进行诊断，但大多数人工智能（AI）模型使用单平面输入，各平面的作用尚未得到充分探索。本研究旨在探讨矢状面、冠状面和横断面膝关节MRI平面在深度学习（DL）模型中对ACL撕裂和半月板撕裂检测的不同影响及相互作用。
使用由1130个训练病例和120个验证病例组成的MRNet数据集来开发TripleMRNet模型。该模型在来自一个、两个或三个平面的图像上进行训练，产生七种组合。本研究通过梯度加权类激活映射（Grad-CAM）提供可解释性分析，系统地比较了这些组合的诊断性能。
对于ACL撕裂检测，三平面模型表现出最高性能，准确率（ACC）为0.925，灵敏度（SEN）为0.944，特异度（SPE）为0.909，F1分数为0.919。冠状面模型的ACC（0.842）、SPE（0.833）和F1分数（0.829）最低。对于半月板撕裂检测，尽管所有七个模型的SEN相似，但三平面模型在ACC（0.783）、SPE（0.824）和F1分数（0.745）方面表现出更好的性能。在这三个指标上，横断面模型仅次于三平面模型，差距很小。相反，矢状面模型表现最差，ACC为0.633，SPE为0.545，F1分数为0.639。
矢状面被证明对检测ACL撕裂最有效，横断面MR图像也显示出显著效用。对于半月板撕裂检测，横断面明显优于其他两个平面，矢状面表现最差。
