• Michail E Klontzas, Evangelia E Vassalou, George A Kakkos, Konstantinos Spanakis, Aristeidis Zibis, Kostas Marias, and Apostolos H Karantanas.
• Department of Medical Imaging, University Hospital of Heraklion, Voutes, 71110, Crete, Greece; Computational BioMedicine Laboratory, Institute of Computer Science, Foundation for Research and Technology (FORTH), N. Plastira 100, 70013, Heraklion, Crete, Greece; Advanced Hybrid Imaging Systems, Institute of Computer Science, Foundation for Research and Technology (FORTH), N. Plastira 100, 70013, Heraklion, Crete, Greece; Department of Radiology, School of Medicine, University of Crete, Voutes Campus, 71003, Heraklion, Greece. Electronic address: miklontzas@ics.forth.gr.
• Injury. 2022 Jun 1; 53 (6): 2035-2040.

PurposeSubchondral insufficiency fractures (SIF) and advanced osteoarthritis (OA) of the knee are usually seen in conjunction with bone marrow lesions (BMLs) and their differentiation may pose a significant diagnostic challenge. We aimed to develop a convolutional neural network (CNN) ensemble which could successfully differentiate between these two entities.Materials And MethodsA total of 212 knees with SIF and 102 knees with advanced OA with BMLs were retrospectively included. Coronal fat suppressed PD-w images were augmented, resized and normalized, reaching a total of 1174 images. Data was used to fine-tune three ImageNet-pretrained CNNs (VGG-16, InceptionV3 and Inception-ResNet-V2). Agreement of at least two networks was recorded as the decision of the network ensemble. Ensemble performance was compared to that of two MSK radiologists on the validation set. Receiver operating characteristics (ROC) curves and the respective areas under the curve (AUC) were used to evaluate human and machine performance.ResultsInceptionV3 achieved the highest AUC (93.68%) and VGG-16 the lowest AUC (82.18%) among individual CNNs. CNN ensemble achieved the highest overall performance with an AUC of 95.97%. The first of the two MSK radiologists achieved a performance similar to the ensemble, reaching an AUC of 91.95%. The second radiologist achieved lower AUC of 82.76% which was lower than both the other specialist and the ensemble (P < 0.001).ConclusionA CNN ensemble was highly accurate in differentiating between SIF and OA, achieving a higher or equal performance to MSK radiologists.Copyright © 2022 Elsevier Ltd. All rights reserved.

15 keywords...

hide…