AUTHOR=Xiongfeng Tang , Yingzhi Li , Xianyue Shen , Meng He , Bo Chen , Deming Guo , Yanguo Qin 

TITLE=Automated detection of knee cystic lesions on magnetic resonance imaging using deep learning

JOURNAL=Frontiers in Medicine

VOLUME=Volume 9 - 2022

YEAR=2022

URL=https://www.frontiersin.org/journals/medicine/articles/10.3389/fmed.2022.928642

DOI=10.3389/fmed.2022.928642

ISSN=2296-858X

ABSTRACT=Background: cystic lesions are frequently revealed in knee joint disease and are usually associated with joint pain and degenerative disorders or acute injury. Therefore, artificial intelligence-assisted cyst detection based on magnetic resonance imaging is one of the effective methods to improve the analysis ability of the whole knee joint. However, a few papers paid attention to it, This research is the first attempt at auto-detection of knee cysts based on deep learning methods.
Methods: In this retrospective study, we collected 282 cysts confirmed subjects in our institution from January to October 2021. We developed a SE-YOLOv5 model based on a self-attention mechanism for knee cyst like lesion detection and distinguish it from knee effusions which also show high light signal in Magnetic Resonance Images. the entire image was resized into 640*640 and then taken as input to the model, the target frame of a position was directly returned to the image. We also add an attention mechanism called the SE module in the model to enhance the contribution of information-rich features in the feature extraction process. accuracy, precision, recall, mean average precision, F1 score, and Frames Per Second (fps) were used to validate the performance of models.
Results: The experimental results show that the method can accurately identify knee MRI images, and auto-detect cyst lesions, not only for obvious cysts but the small and the cyst with inconspicuous contrasts, and the detection speed is greatly improved compared with the comparison algorithm. The SE-YOLO V5 model we presented has the best performance, scoring 0.88 on F1, 0.89 on precision, 0.86 on recall, and all mAP0.5 scores of 0.95, 0.95 for effusion, and 0.94 for the cyst. Meanwhile, the fps for SE-Yolo v5 is 90.9, which means it can handle more images per second. 
Conclusions: This proof-of-principle study examined whether deep learning can detect knee cysts and classify them from knee effusions. Our experiments demonstrated that the classical Yolo V5 and proposed SE-Yolo V5 models could identify cysts with high accuracy.