V Roblot1, Y Giret2, M Bou Antoun3, C Morillot4, X Chassin4, A Cotten5, J Zerbib3, L Fournier6. 1. UMR-S970, Department of Radiology, Hôpital Européen Georges-Pompidou, Assistance Publique-Hôpitaux de Paris, Université Paris-Descartes, 75015 Paris, France. Electronic address: victoire.roblot@aphp.fr. 2. CentraleSupélec, Université Paris Saclay, 91190 Gif-sur-Yvette, France; Foodvisor, 75011 Paris, France. 3. UMR-S970, Department of Radiology, Hôpital Européen Georges-Pompidou, Assistance Publique-Hôpitaux de Paris, Université Paris-Descartes, 75015 Paris, France. 4. CentraleSupélec, Université Paris Saclay, 91190 Gif-sur-Yvette, France. 5. Department of Musculoskeletal Radiology, Lille University Hospital, 59037 Lille, France. 6. UMR-S970, Department of Radiology, Hôpital Européen Georges-Pompidou, Assistance Publique-Hôpitaux de Paris, Université Paris-Descartes, 75015 Paris, France; Laboratoire de Recherche en Imagerie, LRI, PARCC-HEGP, UMR 970, Inserm/université Paris Descartes, Sorbonne-Paris cité, 75015 Paris, France.
Abstract
PURPOSE: The purpose of this study was to build and evaluate a high-performance algorithm to detect and characterize the presence of a meniscus tear on magnetic resonance imaging examination (MRI) of the knee. MATERIAL AND METHODS: An algorithm was trained on a dataset of 1123 MR images of the knee. We separated the main task into three sub-tasks: first to detect the position of both horns, second to detect the presence of a tear, and last to determine the orientation of the tear. An algorithm based on fast-region convolutional neural network (CNN) and faster-region CNN, was developed to classify the tasks. The algorithm was thus used on a test dataset composed of 700 images for external validation. The performance metric was based on area under the curve (AUC) analysis for each task and a final weighted AUC encompassing the three tasks was calculated. RESULTS: The use of our algorithm yielded an AUC of 0.92 for the detection of the position of the two meniscal horns, of 0.94 for the presence of a meniscal tear and of 083 for determining the orientation of the tear, resulting in a final weighted AUC of 0.90. CONCLUSION: We demonstrate that our algorithm based on fast-region CNN is able to detect meniscal tears and is a first step towards developing more end-to-end artificial intelligence-powered diagnostic tools.
PURPOSE: The purpose of this study was to build and evaluate a high-performance algorithm to detect and characterize the presence of a meniscus tear on magnetic resonance imaging examination (MRI) of the knee. MATERIAL AND METHODS: An algorithm was trained on a dataset of 1123 MR images of the knee. We separated the main task into three sub-tasks: first to detect the position of both horns, second to detect the presence of a tear, and last to determine the orientation of the tear. An algorithm based on fast-region convolutional neural network (CNN) and faster-region CNN, was developed to classify the tasks. The algorithm was thus used on a test dataset composed of 700 images for external validation. The performance metric was based on area under the curve (AUC) analysis for each task and a final weighted AUC encompassing the three tasks was calculated. RESULTS: The use of our algorithm yielded an AUC of 0.92 for the detection of the position of the two meniscal horns, of 0.94 for the presence of a meniscal tear and of 083 for determining the orientation of the tear, resulting in a final weighted AUC of 0.90. CONCLUSION: We demonstrate that our algorithm based on fast-region CNN is able to detect meniscal tears and is a first step towards developing more end-to-end artificial intelligence-powered diagnostic tools.
Authors: Gabby B Joseph; Charles E McCulloch; Jae Ho Sohn; Valentina Pedoia; Sharmila Majumdar; Thomas M Link Journal: Skeletal Radiol Date: 2021-11-04 Impact factor: 2.199
Authors: Laura M Fayad; Vishwa S Parekh; Rodrigo de Castro Luna; Charles C Ko; Dharmesh Tank; Jan Fritz; Shivani Ahlawat; Michael A Jacobs Journal: Invest Radiol Date: 2021-06-01 Impact factor: 10.065