gms | German Medical Science

Objectives: Total hip replacements (THR) are one of the most frequently performed surgeries in Germany. Precise planning and subsequent measuring of radiographic hip and pelvis parameters is essential for a satisfactory outcome for the patient. During clinical routine the parameters are measured manually by physicians in a labour-intensive, time-consuming and observer-dependent process. The goal of the study is to develop and validate an algorithm for the automatic and accurate computation of key radiographic parameters using state-of-the-art methods including Artificial Intelligence (AI).

Methods: In a retrospective study, AP pelvic radiographs of 100 patients (hips with/without THR: 103/97, female/male: 59/41) were measured twice by an experienced physician and compared with the results of an AI-based algorithm. Repeated measurements of the physician were used to analyze the intra-rater-reliability. A pipeline was developed containing two trained Neural Networks to first locate anatomic regions of interest in AP pelvis X-rays and to detect specific anatomical landmarks (e.g., pelvic tear drop, lesser trochanter). The resultinginformation was used for the fully automated determination of the following radiographic parameters in accordance to the scientific literature [1]: Femoral Offset (FO), Leg Length Difference (LLD), Caput-Collum-Diaphyseal Angle (CCD) and the inclination of two reference lines, the Inter-Teardrop Line (ITL) and the Biischial Line (BL). To evaluate the performance of the AI algorithm, mean error, standard deviation and inter-rater-reliability were assessed using single measure Intraclass Correlation Coefficient (ICC, absolute agreement). ICC > 0.75 were considered excellent [2].

Results and conclusion: The ICC values for the intra-rater-reliability (0.92 – 0.99; LLD andBL) show excellent agreement for multiple measurements from the same physician. The AI algorithm was able to assess all parameters in all 100 images, reflected by excellent ICC values: 0.89 (FO), 0.82 (LLD), 0.90 (CCD) and for the inclination of the reference lines 0.94 (ITL), 0.88 (BL) compared to the human rater. The mean errors and standard deviations for the parameters were -2.3 ± 2.7 mm (FO), 0.3 ± 4.3 mm (LLD), 0.4 ± 3.7° (CCD) and for the reference lines -0.2 ± 0.9° (ITL), 0.0 ± 1.5° (BL). The excellent agreement between the measurements done by a physician and the fully automated algorithm suggests high reliability and accuracy. Therefore, the algorithm could support physicians in their daily routine in the assessment of hip and pelvis parameters or in the analysis of large data sets for research purposes.

Figure 1 [Fig. 1]