A fully automatic system to assess foot collapse on lateral weight-bearing foot radiographs: A pilot study

Abstract

Background: Foot collapse is primarily diagnosed and monitored using lateral weight-bearing foot x-ray images. There are several well-validated measurements which aid assessment. However, these are subject to inter- and intra-user variability.

Objective: To develop and validate a software system for the fully automatic assessment of radiographic changes associated with foot collapse; automatically generating measurements for calcaneal tilt, cuboid height and Meary's angle.

Methods: This retrospective study was approved by the Health Research Authority (IRAS 244852). The system was developed using lateral weight-bearing foot x-ray images, and evaluated against manual measurements from five clinical experts. The system has two main components: (i) a Random Forest-based point-finder to outline the bones of interest; and (ii) a geometry-calculator to generate the measurements based on the point positions from the point-finder. The performance of the point-finder was assessed using the point-to-point error (i.e. the mean absolute distance between each found point and the equivalent ground truth point, averaged over all points per image). For assessing the performance of the geometry-calculator, linear mixed models were fitted to estimate clinical inter-observer agreement and to compare the performance of the software system to that of the clinical experts.

Results: A total of 200 images were collected from 79 subjects (mean age: 56.4 years ±12.9 SD, 30/49 females/males). There was good agreement among all clinical experts with intraclass correlation estimates between 0.78 and 0.86. The point-finder achieved a median point-to-point error of 2.2 mm. There was no significant difference between the clinical and automatically generated measurements using the point-finder points, suggesting that the fully automatically obtained measurements are in agreement with the manually obtained measurements.

Conclusions: The proposed system can be used to support and automate radiographic image assessment for diagnosing and managing foot collapse, saving clinician time, and improving patient outcomes.

Keywords: Charcot foot; Clinical decision support system; Diabetes; Landmark localization; Radiomics.

Fig. 1

Flow diagram showing the number of included participants and corresponding number of x-ray images (IQR = interquartile range).

Fig. 2

Visualisation of the measurements used in this study: calcaneal tilt (blue), cuboid height (yellow) and Meary's angle (red); (CN = Charcot neuroarthropathy).

Fig. 3

Examples of varying degrees of foot collapse based on Meary's angle (MMA = mean Meary's angle over all five clinical experts). The same images are classified as normal, mild, severe and severe based on clinical review. All radiographs have been cropped for better visualisation.

Fig. 4

Annotation examples with 61 points (manually placed): (a) showing the 61 point positions and point indices for a mild case based on Meary's angle (MMA: 11.0°) with poor collimation. (b) showing the 61 point positions for a mild case based on Meary's angle (MMA: 6.3°) which shows advanced midfoot collapse. In this case, the 1st metatarsal has subluxed superiorly causing Meary's angle to be almost normal. This is a severe case based on clinical review. All radiographs have been cropped for better visualisation (MMA = mean Meary's angle over all five clinical experts).

Fig. 5

CT (a), CH (c) and MA (e) are manual measurement values per severity class based on Meary's angle: normal (cyan), mild (purple), moderate (orange), severe (magenta). CT (b), CH (d) and MA (f) are manual measurement values per severity class based on clinical review: normal (cyan), mild (orange), severe (magenta). All five manual measurements per image are shown and joined with a line for better visibility. Automatic results (black) are largely in agreement with manual measurements (i.e. black dots are within spread of manual measurements per image). The images in all plots are sorted by the mean over the respective manual measurements.

Fig. 6

Fully automatic point-finder performance for locating the 61 points as in Fig. 4 in all 200 images: (a, b) showing the performance for both the RFRV-CLM and the SCN methods; all results were obtained using the same data and 10-fold cross-validation experiments. (c, d) showing the RFRV-CLM performance for different severity classes with severity defined based on the average manual Meary's angle. (e, f) showing the RFRV-CLM performance for different severity classes with severity defined based on clinical review. The severity grouping was applied retrospectively to the point-finder results; the point positions of all groups were obtained using the same 10-fold cross-validation point-finder systems.