Hämäläinen Mathias, Sormaala Markus, Kaseva Tuomas, Salli Eero, Savolainen Sauli, Kangasniemi Marko
Department of Radiology, HUS Diagnostic Center, Helsinki University Hospital and University of Helsinki, Jorvin Sairaala, Karvasmäentie 8, Espoo, 02740, Finland.
Department of Radiology, HUS Diagnostic Center, Helsinki University Hospital and University of Helsinki, Haartmaninkatu 4, P.O. Box 340, Helsinki, 00290, Finland.
BMC Musculoskelet Disord. 2025 Aug 7;26(1):760. doi: 10.1186/s12891-025-09011-1.
To investigate feasibility of a method which combines segmenting convolutional neural networks (CNN) for the automated detection of ganglion cysts in 2D MRI of the wrist. The study serves as proof-of-concept, demonstrating a method to decrease false positives and offering an efficient solution for ganglia detection.
We retrospectively analyzed 58 MRI studies with wrist ganglia, each including 2D axial, sagittal, and coronal series. Manual segmentations were performed by a radiologist and used to train CNNs for automatic segmentation of each orthogonal series. Predictions were fused into a single 3D volume using a proposed prediction fusion method. Performance was evaluated over all studies using six-fold cross-validation, comparing method variations with metrics including true positive rate, number of false positives, and F-score metrics.
The proposed method reached mean TPR of 0.57, mean FP of 0.4 and mean F-score of 0.53. Fusion of series predictions decreased the number of false positives significantly but also decreased TPR values.
CNNs can detect ganglion cysts in wrist MRI. The number of false positives can be decreased by a method of prediction fusion from multiple CNNs.
研究一种结合分割卷积神经网络（CNN）在腕部二维磁共振成像（MRI）中自动检测腱鞘囊肿方法的可行性。该研究作为概念验证，展示了一种减少假阳性的方法，并为腱鞘囊肿检测提供了一种有效的解决方案。
我们回顾性分析了58例患有腕部腱鞘囊肿的MRI研究，每项研究包括二维轴向、矢状和冠状序列。由一名放射科医生进行手动分割，并用于训练CNN对每个正交序列进行自动分割。使用一种提出的预测融合方法将预测结果融合为单个三维体积。使用六折交叉验证对所有研究的性能进行评估，将方法变体与包括真阳性率、假阳性数量和F分数指标在内的指标进行比较。
所提出的方法达到了平均真阳性率0.57、平均假阳性0.4和平均F分数0.53。序列预测的融合显著减少了假阳性数量，但也降低了真阳性率值。
CNN可以在腕部MRI中检测腱鞘囊肿。通过来自多个CNN的预测融合方法可以减少假阳性数量。
