Pes planus, or flatfoot, is a prevalent condition characterized by inward collapse of the arch along the inner side of the foot (Lalevée et al., 2022; Mann & Thompson, 1985). It is commonly categorized into two types: pediatric pes planus and adult-acquired pes planus (AAPP). Pediatric pes planus is also categorized into two groups: flexible pes planus and painful pes planus. The flexible type is typically characterized by hypermobility of the subtalar joint, and development of the medial arch normally occurs naturally before adulthood (Gül et al., 2023; Ueki et al., 2019). However, painful pes planus in children and/or adolescents is caused mostly by the talocalcaneal coalition (Guduri & Dreyer, 2019). AAPP is acknowledged as a source of discomfort and impairment in adulthood, primarily resulting from posterior tibial tendon deficiency (PTTD) (Smyth et al., 2017).
Tarsal coalition is when two or more tarsal bones fuse together incorrectly, which can lead to painful pes planus in adolescents. Tarsal coalitions are classified into fibrous, cartilaginous, and osseous coalitions based on the abnormal bridge morphology (Catanzano Jr et al., 2023). The exact prevalence of tarsal coalition is uncertain; estimates vary from less than 1% to around 1–2% of the population (Bernasconi et al., 2017).
The prevalence of AAPP ranges from 5–15% in the community. Approximately 1 out of every 100 people in the overall population seeks medical attention for the pain resulting from a flat foot. The prevalence of asymptomatic AAPP, which is also quite prevalent, is not fully known (Ferciot, 1972; Ling & Lui, 2017).
PTTD, the most common reason for AAPP, can be attributed to various factors, including age, obesity, ligamentous laxity, and trauma. There is a watershed area behind the medial malleolus that typically experiences degeneration as a result of a sudden alteration in the orientation of the tibialis posterior tendon. PTTD results in a disruption of force balance (Rhim et al., 2022; Smyth et al., 2017; Ueki et al., 2019). Peroneus brevis tendon causes forefoot abduction and the Achilles tendon causes a valgus deformity of the ankle. Subsequently, spring tendon weakness ensues, leading to collapse of the medial arch (Cifuentes-De la Portilla et al., 2019; Kaye & Jahss, 1991).
Pes planus can be diagnosed by both clinical examination and radiographic methods. The primary radiological approach for investigating pes planus is a full-weight lateral foot X-ray in the stance phase (Fadle et al., 2023; Gould, 1982). Assessment of AAPP involves measurements on lateral weight-bearing X-rays. These measurements include the talo-first metatarsal angle (Meary’s angle), calcaneal pitch angle, and height between the medial cuneiform and the floor. However, using these parameters can be time consuming and may lead to inconsistent results between different observers (Bock et al., 2018).
The potential of computer-based, automated systems for disease detection from X-ray images is extremely promising. X-ray images are considered the primary method for the evaluation of pes planus. There are many studies in the literature on pes planus detection from X-ray images with machine learning. There are two basic approaches to pes planus. The first is to determine whether it exists by classifying the images. The second is to calculate the angles used to detect pes planus from the patient's foot X-ray images. Image interpretation by computer-aided systems equipped with artificial intelligence can be more objective, faster, and more accurate than human interpretation (Stotter et al., 2023). Therefore, computer-based automatic detection of pes planus is an important research topic in the literature.
Jian et al. used the Canny edge detection algorithm to find the arch angle required to detect pes planus. Pes planus was detected according to the principle of obtaining the arch angle by identifying key points. They also used a Gaussian filter to remove noise in the images (Jian et al., 2014). Image processing-based applications significantly affect image contrast performance. Yang et al. identified the landmarks necessary to find the arch angle. They used mutual information (MI) for this. They looked at the angle difference between the two images using the reference image and the template image. The rotation of the template image relative to the reference image contains information about the arch angle. They reported a 96% accuracy rate (Yang et al., 2015).
Kao et al. counted the white pixels in the image to identify the calcaneal and fifth metatarsal bones. They found the arch angle by determining key points for the calcaneal and fifth metatarsals through binary images. They stated that 73% of all cases were detected correctly (Kao et al., 2018).
Machine learning algorithms are also used effectively to detect pes planus. Skwirczyński et al. proposed a method based on a random forest (RF). The key points required to find the calcaneal angle, mean angle, and talar angle were scanned in the image with a RF. Each pixel was assigned a weight value depending on whether it was a key point or not. The angles between key points were calculated by classification according to their weight values (Skwirczyński et al., 2019).
Nitris et al. proposed feature extraction based on ResNet50. They suggested a new convolutional neural network (CNN) model that includes Adam optimization. They divided the image into three regions and detected key points separately from each region (Nitris et al., 2019).
According to Gül et al. (2023), pes planus detection was conducted based on the classification of X-ray images. The performance of different CNN models was examined. Display features were extracted with MobilenetV2, which provides the highest performance. Then dimension reduction was carried out by feature selection. Classification was performed with a support vector machine (SVM) and 95.14% accuracy was achieved.
In the present study, we proposed a new fusion model for computer-aided detection of pes planus. This fusion model consists of a combination of the CNN network and the vision transformer (ViT) network. In this way, we benefited from the advantages of both networks together. A robust pes planus detection system was developed through ensemble learning. The methodology is explained in the second part of the study, the experimental results are presented in the third part, the discussion is set out in the fourth part, and the conclusions are given in the last part.