A Novel Deep Transfer Learning-Based Approach for Automated Pes Planus Diagnosis Using X-ray Image

Abstract

Pes planus, colloquially known as flatfoot, is a deformity defined as the collapse, flattening or loss of the medial longitudinal arch of the foot. The first standard radiographic examination for diagnosing pes planus involves lateral and dorsoplantar weight-bearing radiographs. Recently, many artificial intelligence-based computer-aided diagnosis (CAD) systems and models have been developed for the detection of various diseases from radiological images. However, to the best of our knowledge, no model and system has been proposed in the literature for automated pes planus diagnosis using X-ray images. This study presents a novel deep learning-based model for automated pes planus diagnosis using X-ray images, a first in the literature. To perform this study, a new pes planus dataset consisting of weight-bearing X-ray images was collected and labeled by specialist radiologists. In the preprocessing stage, the number of X-ray images was augmented and then divided into 4 and 16 patches, respectively in a pyramidal fashion. Thus, a total of 21 images are obtained for each image, including 20 patches and one original image. These 21 images were then fed to the pre-trained MobileNetV2 and 21,000 features were extracted from the Logits layer. Among the extracted deep features, the most important 1312 features were selected using the proposed iterative ReliefF algorithm, and then classified with support vector machine (SVM). The proposed deep learning-based framework achieved 95.14% accuracy using 10-fold cross validation. The results demonstrate that our transfer learning-based model can be used as an auxiliary tool for diagnosing pes planus in clinical practice.

Keywords: X-ray image classification; deep learning; iterative ReliefF; pes planus; pyramidal feature extraction; transfer learning.

Figure 1

Randomly selected X-ray images from the labeled pes planus dataset.

Figure 2

The framework of the proposed method.

Figure 3

The process of images dividing into patches in a pyramidal fashion.

Figure 4

The main blocks of CNN architecture.

Figure 5

The feature selection process. (a) Weight value calculated by ReliefF algorithm for 21,000 features using 10 nearest neighbors. (b) The weight of the 1312 most important features selected from 21,000 features by the proposed iterative ReliefF algorithm.

Figure 6

Confusion matrix of the proposed pyramidal-MobileNetV2-cubic SVM-based model.