gms | German Medical Science

Objectives: Within surgical treatment options for knee osteoarthritis such as arthroplasty and osteotomy, a comprehensive analysis of the leg alignment is paramount. A deep learning (DL) model that performs automated analysis of the leg alignment on x-rays could accelerate the process currently performed by orthopedic surgeons (OS) and increase reliability of preoperative planning. The purpose of this study was to develop a DL model for automated and accurate assessment of the leg alignment on anterior posterior (ap) hip-knee-ankle (HKA) x-rays and compare the performance to OS.

Methods: 594 patients (mean age 41.1±13.2 years, 182 female, 388 left side), who underwent osteotomy at the authors' institution were retrospectively enrolled. On a.p. HKA radiographs (Figure 1a [Fig. 1]), alignment analysis and placement of landmarks was performed (Figure 1b - red [Fig. 1]) by two OS (OS1 and OS2), serving as ground truth. Measurements included the mechanical femorotibial angle (mFA-mTA), lateral distal femoral angle (mLDFA), medial proximal tibia angle (mMPTA), lateral distal talus angle (mLDTA), joint line convergence angle (JLCA) and anatomical angle (AMA). The data set was split 60% (n=399) / 10% (n=59) / 30% (n=136) for training, validation, hold-out testing. Twelve networks - each specialized on an anatomical region (Figure 1b - green [Fig. 1]) - were synthesized and angles were calculated (Figure 1c [Fig. 1]). The model was based on a COCO pretrained Mask-R CNN-ResNeXt-101 implemented in PyTorch. The mean difference of the individual angles between the DL model and the ground truth was measured in the hold-out test set and compared to the mean difference of OS1 and OS2 to evaluate the performance of the DL model.

Results and conclusion: Excellent agreement between the predicted landmarks by the DL model and the annotated landmarks (Figure 1b [Fig. 1]) allowed for an accurate calculation of angles. The mean difference between the DL model and the ground truth was 0.1°±0.1° for mFA-mTA, 0.7°±0.7° for mLPFA, 1.0°± 0.8° for mLDFA,0.6°±0.6° for mMPTA, 0.9°±0.9° for mLDTA, 0.5°± 0.8° for JLCA and 0.3°±0.5° for AMA. In comparison, the mean difference between OS1 and OS2 was 0.3°±0.2° for mFA-mTA, 1.0°±1.1° for mLDFA, 1.1°±2.0° for mMPTA, 1.4°±1.8° for mLDTA, 0.9°±1.2° for JLCA and 0.3°±0.3° for AMA. The DL model outperformed the OS in the time required for the analysis (22,4±0,5s vs. 91,0±10,0s).

The developed DL model combining landmark detection with segmentation and object detection allowed for an accurate assessment of the leg alignment on a.p. HKA radiographs with a performance comparable to OS, yet in a substantially shorter time. A model such as this could accelerate preoperative planning, as well as significantly increase its accuracy and reliability.