PREDICTING MECHANICAL LEG ALIGNMENT FROM KNEE X-RAY IMAGES BY UTILIZING DEEP-LEARNING NEURAL NETWORKS

Several studies have linked leg malalignment to early onset and progression of knee OA. Detecting abnormalities in mechanical alignment before radiographic evidence of knee OA is present, might increase the rate of early detection. Hence, conservative joint preserving interventions or even osteotomies could reduce the need or delay full joint replacement. In addition, medical trials often face high dropout rates, consequently, AI based support could improve the recruitment process. Detecting mechanical leg malalignment: 1) before the malalignment is immediately obvious to the clinician and 2) excluding/including patients recruited for medical trials. This study utilized a dataset of 8,878 digital radiographs ((I) 6,181 AP/PA full-leg x-rays, (II) 2,292 AP/PA knee x-rays with SynaFlexer positioning frame, (III) 405 AP/PA knee x-rays captured without SynaFlexer). The images were obtained from a diverse range of sources, including the OAI study, the University for Continuing Education Krems (AT), and five additional sites in Austria and the US. All knee images in the dataset were accompanied by their corresponding full-leg images, obtained on the same day. To determine the ground-truth hip-knee-ankle angle (HKA), the full-leg images were processed using IB Lab LAMA (ImageBiopsyB Lab, Vienna, Austria), a commercial software that automatically determines clinical measurements from full-leg radiographs. These images were used to train the AI algorithm. The algorithm produced binary outputs for HKA threshold values of 5° (clinically relevant) and 7.5° (clinical trial relevant). The algorithm reached MAE rates of 0.98° (cropped full-leg), 1.56° (SynaFlexer knee) and 2.10° (non-SynaFlexer knee). Summarizing all image modalities, the overall performance for the 5° threshold was 74-86% Sensitivity and 79-90% Specificity. Following, 7.5° threshold performance reached 81-92% Sensitivity and 85-92% Specificity. While this study primarily focused on absolute HKA threshold values of 5° and 7.5° for varus/valgus classification, this threshold determination process is highly adaptable to accommodate specific use cases, such as the desired angle threshold for classification or the preferred balance between sensitivity and specificity. In addition, overestimated HKA have an elevated angle above the normal range and could proceed towards or above mentioned thresholds. Chen K. PhD funding “LowerAustria Society for Scientific Research” (LSC20-020). Salzlechner C. and Lepenik C.