Artificial intelligence-based automatic assessment of lower limb torsion on MRI

Abstract

Abnormal torsion of the lower limbs may adversely affect joint health. This study developed and validated a deep learning-based method for automatic measurement of femoral and tibial torsion on MRI. Axial T2-weighted sequences acquired of the hips, knees, and ankles of 93 patients (mean age, 13 ± 5 years; 52 males) were included and allocated to training (n = 60), validation (n = 9), and test sets (n = 24). A U-net convolutional neural network was trained to segment both femur and tibia, identify osseous anatomic landmarks, define pertinent reference lines, and quantify femoral and tibial torsion. Manual measurements by two radiologists provided the reference standard. Inter-reader comparisons were performed using repeated-measures ANOVA, Pearson's r, and the intraclass correlation coefficient (ICC). Mean Sørensen-Dice coefficients for segmentation accuracy ranged between 0.89 and 0.93 and erroneous segmentations were scarce. Ranges of torsion as measured by both readers and the algorithm on the same axial image were 15.8°-18.0° (femur) and 33.9°-35.2° (tibia). Correlation coefficients (ranges, .968 ≤ r ≤ .984 [femur]; .867 ≤ r ≤ .904 [tibia]) and ICCs (ranges, .963 ≤ ICC ≤ .974 [femur]; .867 ≤ ICC ≤ .894 [tibia]) indicated excellent inter-reader agreement. Algorithm-based analysis was faster than manual analysis (7 vs 207 vs 230 s, p < .001). In conclusion, fully automatic measurement of torsional alignment is accurate, reliable, and sufficiently fast for clinical workflows.

Figure 1

Representative cases that challenged correct segmentation and determination of reference lines. (a) Motion artefacts during the pelvic MRI scan (a₁) caused smearing of the images at the level of the femoral head (a₂) and neck (a₃), erroneous delineation of the femoral head contour (a₄), and incorrect identification of the femoral neck axis through the greater trochanter (a₅). (b) Motion artefacts during the MRI scan of the knee (b₁) created an artificial protrusion of the medial femur (b₂) that did not negatively affect the reference line (b₃). (c, d) Prominent and variable epiphyseal plates at the tibia (c) or femur (d) were associated with inaccurate segmentations (block arrows in c₂, d₂) that caused misplacement of the reference line (c₃) or not (d₃). (e) Post-surgical changes and fat graft interposition at the anteromedial tibia (°) after physiolysis (e₁) caused inaccurate segmentation (e₂) but did not negatively affect the reference line (e₃). (f) At the hip, immature bone characterized by the cartilaginous greater trochanter and femoral neck (* in f₁) is segmented incorrectly (f₂) and the femoral neck axis is identified too caudal (f₃). (g) Similarly, cartilaginous femoral and tibial condyles (* in g₁) are characzterized by incorrect segmentations and reference lines (g₂, g₃). Inset coronal images framed in blue indicate the height of the axial slices. Patient age and gender: 9 years and male (a, c), 11 years and male (b, e), 12 years and male (d), 5 years and male (g).

Figure 2

Manual reference measurements to determine femoral and tibial torsion at the levels of the hip, knee, and anke. Anatomic landmarks were used to define the reference lines at the hip in line with the method suggested by Lee (PF proximal femur). The reference lines at the knee were delineated as the distal femoral reference line (DF distal femur) and as the proximal tibial reference line (PT proximal tibia), while the reference line at the ankle was determined using the ellipses method (DT distal tibia). The circles indicate the (superimposed) femoral head and ellipses along the medial malleolus and fibular notch, while dotted lines visualize the horizontal reference lines. Schematics of femur (yellow), tibia (purple), and fibula (light blue) on the right indicate the levels of the axial images. 12-year-old female. Please refer to Supplementary Figure 3 for a visualization of the other manual reference methods to determine femoral torsion (according to the Reikeras, Tomczak, and Murphy methods) and tibial torsion (according to the bimalleolar and talus methods).

Figure 3

Schematic visualization of the U-net convolutional neural network used for automatic segmentations of femur (yellow), tibia (purple), and fibula (light blue). Global information were compressed to a more compact global representation. Local information was preserved by skip connections which passed the more detailed images directly to the decoder path. Up-sampling was performed by bilinear interpolation followed by 1 × 1x1 convolutions. The rectified linear unit (ReLU) was applied in all layers except the last, where a Softmax layer determined each voxels’ class probabilities. 3D convolutions aimed to reflect the high dimensionality of the MRI data. Instance normalization was chosen due to small batch size.

Figure 4

Algorithm-based identification of proximal and distal femoral and tibial reference lines to determine lower limb torsion. (a) At the hip, the femoral head centre was identified and used to define the femoral neck axis that served as the proximal femoral reference line. (b, c) Around the knee, the most posterior extensions of the medial and lateral femoral and tibial condyles were identified and connected as the distal femoral (b) and proximal tibial reference lines (c). (d) At the ankle, the tibial and fibular centroids were connected as the distal tibial reference line. RL reference line. MR images are not to scale.

Figure 5

Visualization of the algorithm-based identification of the reference lines. Indicated are the original MR images (a₁-d₁), their segmentation outlines on selected axial images (a₂-d₂), and the corresponding reference lines (a₃-d₃) as determined by the algorithm. Proximal femur (a), distal femur (b), proximal tibia (c), and distal tibia (d). Reference lines and points are blue. For algorithmic details please refer to Fig. 4. Femur (yellow), tibia (purple), and fibula (light blue). 19-year-old female.