Enhanced diagnosis of pes planus and pes cavus using deep learning-based segmentation of weight-bearing lateral foot radiographs: a comparative observer study

A weight-bearing lateral radiograph (WBLR) of the foot is a gold standard for diagnosing adult-acquired flatfoot deformity. However, it is difficult to measure the major axis of bones in WBLR without using auxiliary lines. Herein, we develop semantic segmentation with a deep learning model (DLm) on the WBLR of the foot for enhanced diagnosis of pes planus and pes cavus. We used 300 consecutive WBLRs from young Korean males. The semantic segmentation model was developed based on U2-Net. An expert orthopedic surgeon manually labeled ground truths. We used 200 radiographs for training, 100 for internal validation, and two external datasets for external validation. The model was trained using a hybrid loss function, combining Dice Loss and boundary-based loss, to enhance both overall segmentation accuracy and precise delineation of boundary regions between pes planus and pes cavus. Angle measurement errors with minimum moment of inertia (MMI) and ellipsoidal fitting (EF) based on the segmentation results were evaluated. The DLm exhibited better results than human observers. For internal validation, the absolute angle errors of the DLm using MMI and EF were 0.92 +/- 1.32 degrees and 1.34 +/- 2.07 degrees, respectively. In external validation, these errors were 1.17 +/- 1.60 degrees and 1.60 +/- 2.42 degrees for AMC's dataset, and 1.23 +/- 1.39 degrees and 1.68 +/- 1.98 degrees for the LERA dataset, respectively. The DLm showed higher overall diagnostic accuracy than human observers in identifying flatfoot angles, regardless of the measurement methods. The absolute angle errors and diagnostic accuracy of the developed DLm are superior to those of the three human observers. Furthermore, when comparing the angle measurement methods within the DLm, the MMI method proves to be more accurate than EF. Finally, the proposed deep learning model, particularly with the implementation of the U2-Net demonstrates enhanced boundary segmentation and achieves sufficient external validation results, affirming its applicability in the real clinical setting.