Automated diagnosis of flatfoot using cascaded convolutional neural network for angle measurements in weight-bearing lateral radiographs.

OBJECTIVES: Diagnosis of flatfoot using a radiograph is subject to intra- and inter-observer variabilities. Here, we developed a cascade convolutional neural network (CNN)-based deep learning model (DLM) for an automated angle measurement for flatfoot diagnosis using landmark detection. METHODS: We used 1200 weight-bearing lateral foot radiographs from young adult Korean males for the model development. An experienced orthopedic surgeon identified 22 radiographic landmarks and measured three angles for flatfoot diagnosis that served as the ground truth (GT). Another orthopedic surgeon (OS) and a general physician (GP) independently identified the landmarks of the test dataset and measured the angles using the same method. External validation was performed using 100 and 17 radiographs acquired from a tertiary referral center and a public database, respectively. RESULTS: The DLM showed smaller absolute average errors from the GT for the three angle measurements for flatfoot diagnosis compared with both human observers. Under the guidance of the DLM, the average errors of observers OS and GP decreased from 2.35° ± 3.01° to 1.55° ± 2.09° and from 1.99° ± 2.76° to 1.56° ± 2.19°, respectively (both p < 0.001). The total measurement time decreased from 195 to 135 min in observer OS and from 205 to 155 min in observer GP. The absolute average errors of the DLM in the external validation sets were similar or superior to those of human observers in the original test dataset. CONCLUSIONS: Our CNN model had significantly better accuracy and reliability than human observers in diagnosing flatfoot, and notably improved the accuracy and reliability of human observers. KEY POINTS: • Development of deep learning model (DLM) that allows automated angle measurements for landmark detection based on 1200 weight-bearing lateral radiographs for diagnosing flatfoot. • Our DLM showed smaller absolute average errors for flatfoot diagnosis compared with two human observers. • Under the guidance of the model, the average errors of two human observers decreased and total measurement time also decreased from 195 to 135 min and from 205 to 155 min.

Automated landmark identification for diagnosis of the deformity using a cascade convolutional neural network (FlatNet) on weight-bearing lateral radiographs of the foot.

Landmark detection in flatfoot radiographs is crucial in analyzing foot deformity. Here, we evaluated the accuracy and efficiency of the automated identification of flatfoot landmarks using a newly developed cascade convolutional neural network (CNN) algorithm, Flatfoot Landmarks AnnoTating Network (FlatNet). A total of 1200 consecutive weight-bearing lateral radiographs of the foot were acquired. The first 1050 radiographs were used as the training and tuning, and the following 150 radiographs were used as the test sets, respectively. An expert orthopedic surgeon (A) manually labeled ground truths for twenty-five anatomical landmarks. Two orthopedic surgeons (A and B, each with eight years of clinical experience) and a general physician (GP) independently identified the landmarks of the test sets using the same method. After two weeks, observers B and GP independently identified the landmarks once again using the developed deep learning CNN model (DLm). The X- and Y-coordinates and the mean absolute distance were evaluated. The average differences (mm) from the ground truth were 0.60 ± 0.57, 1.37 ± 1.28, and 1.05 ± 1.23 for the X-coordinate, and 0.46 ± 0.59, 0.97 ± 0.98, and 0.73 ± 0.90 for the Y-coordinate in DLm, B, and GP, respectively. The average differences (mm) from the ground truth were 0.84 ± 0.73, 1.90 ± 1.34, and 1.42 ± 1.40 for the absolute distance in DLm, B, and GP, respectively. Under the guidance of the DLm, the overall differences (mm) from the ground truth were enhanced to 0.87 ± 1.21, 0.69 ± 0.74, and 1.24 ± 1.31 for the X-coordinate, Y-coordinate, and absolute distance, respectively, for observer B. The differences were also enhanced to 0.74 ± 0.73, 0.57 ± 0.63, and 1.04 ± 0.85 for observer GP. The newly developed FlatNet exhibited better accuracy and reliability than the observers. Furthermore, under the FlatNet guidance, the accuracy and reliability of the human observers generally improved.

Enhancement of evaluating flatfoot on a weight-bearing lateral radiograph of the foot with U-Net based semantic segmentation on the long axis of tarsal and metatarsal bones in an active learning manner.

Robust labeling for semantic segmentation in radiographs is labor-intensive. No study has evaluated flatfoot-related deformities using semantic segmentation with U-Net on weight-bearing lateral radiographs. Here, we evaluated the robustness, accuracy enhancement, and efficiency of automated measurements for flatfoot-related angles using semantic segmentation in an active learning manner. A total of 300 consecutive weight-bearing lateral radiographs of the foot were acquired. The first 100 radiographs were used as the test set, and the following 200 radiographs were used as the training and validation sets, respectively. An expert orthopedic surgeon manually labeled ground truths. U-Net was used for model training. The Dice similarity coefficient (DSC) and Hausdorff distance (HD) were used to evaluate the segmentation results. In addition, angle measurement errors with a minimum moment of inertia (MMI) and ellipsoidal fitting (EF) based on the segmentation results were compared between active learning and learning with a pooled dataset. The mean values of DSC, HD, MMI, and EF of the average of all bones were 0.967, 1.274 mm, 0.792°, and 1.147° in active learning, and 0.964, 1.292 mm, 0.828°, and 1.186° in learning with a pooled dataset, respectively. The mean DSC and HD were significantly better in active learning than in learning with a pooled dataset. Labeling of all bones required 0.82 min in active learning and 0.88 min in learning with a pooled dataset. The accuracy and angle errors generally converged in both learning. However, the accuracies based on DSC and HD were significantly better in active learning. Moreover, active learning took less time for labeling, suggesting that active learning could be an accurate and efficient learning strategy for developing flatfoot classifiers based on semantic segmentation.