High accuracy in lower limb alignment analysis using convolutional neural networks, with improvements needed for joint-level metrics.

Objective: Evaluation of long-leg standing radiographs (LSR) is a standardised procedure for analysis of primary or secondary deformities of the lower limbs. Deep-learning convolutional neural networks (CNN) offer the potential to enhance radiological measurement by increasing reproducibility and accuracy. This study aims to evaluate the measurement accuracy of an automated CNN-based planning tool (mediCAD® 7.0; mediCAD Hectec GmbH) of lower limb deformities.

Methods: In a retrospective single-centre study, 164 pre- and postoperative bilateral LSRs with uni- or bilateral posttraumatic knee arthritis undergoing total knee arthroplasty (TKA) were enroled. Alignment parameters relevant to knee arthroplasty and deformity correction were analysed independently by two observers and a CNN. The intraclass correlation coefficient (ICC) was used to evaluate the accuracy between observers and the CNN, which was further evaluated using absolute deviations, limits of agreement (LoA) and root mean square error (RMSE).

Results: CNN evaluation demonstrated high consistency in measuring leg length (ICC > 0.99) and overall lower limb alignment measures of mechanical tibio-femoral angle (mTFA) (ICC > 0.97; RMSE < 1.1°). The mean absolute difference between angular measurements were low for overall lower limb alignment (mTFA 0.49-0.61°) and high for specific joint angles (aMPFA 3.86-4.50°). Accuracy at specific joint angles like the mechanical proximal tibial angle (MPTA) and the mechanical lateral distal femur angle (mLDFA) varied between lower limbs with deformity, with and without TKA with greatest difference for TKA (ICC 0.22-0.85; RMSE 1.72-3.65°).

Conclusion: Excellent accuracy was observed between manual and automated measurements for overall alignment and leg length, but joint-level metrics need further improvement especially in case of TKA similar to other existing algorithms. Despite the observed deviations, the time-efficient nature of the algorithm improves the efficiency of the preoperative planning process.

Level Of Evidence: IV.