Key-Exchange Convolutional Auto-Encoder for Data Augmentation in Early Knee Osteoarthritis Detection

TL;DR

This work tackles the limited annotated data problem in early KOA detection by introducing a Key-Exchange Convolutional Auto-Encoder (KECAE) that generates diverse, clinically valid augmented X-ray images. KECAE uses a dual-input, dual-encoder architecture to separate and swap key pathological features (e.g., JSN and osteophytes) between healthy and KOA images, guided by a hybrid loss that combines reconstruction, discriminator-based supervision, and Fisher LDA-based feature separation. Augmented data improve KOA classification across multiple architectures, with notable gains (e.g., +1.98% for VGG-11 and +0.68% for a transformer baseline) and strong clinical validation showing realistic, interpretable outputs (average accuracy 86.33%, SRS 4.13, kappa 0.78). This approach demonstrates a practical, physiology-aware data augmentation strategy to bolster early KOA detection when labeled data are scarce.

Abstract

Knee Osteoarthritis (KOA) is a common musculoskeletal condition that significantly affects mobility and quality of life, particularly in elderly populations. However, training deep learning models for early KOA classification is often hampered by the limited availability of annotated medical datasets, owing to the high costs and labour-intensive nature of data labelling. Traditional data augmentation techniques, while useful, rely on simple transformations and fail to introduce sufficient diversity into the dataset. To address these challenges, we propose the Key-Exchange Convolutional Auto-Encoder (KECAE) as an innovative Artificial Intelligence (AI)-based data augmentation strategy for early KOA classification. Our model employs a convolutional autoencoder with a novel key-exchange mechanism that generates synthetic images by selectively exchanging key pathological features between X-ray images, which not only diversifies the dataset but also ensures the clinical validity of the augmented data. A hybrid loss function is introduced to supervise feature learning and reconstruction, integrating multiple components, including reconstruction, supervision, and feature separation losses. Experimental results demonstrate that the KECAE-generated data significantly improve the performance of KOA classification models, with accuracy gains of up to 1.98% across various standard and state-of-the-art architectures. Furthermore, a clinical validation study involving expert radiologists confirms the anatomical plausibility and diagnostic realism of the synthetic outputs. These findings highlight the potential of KECAE as a robust tool for augmenting medical datasets in early KOA detection.

Figures (5)

• Figure 1: The flowchart of the proposed KECAE model. Black, blue, and red arrows denote the overall data flow, non-key features, and key features, respectively. The symbol $\oplus$ represents element-wise addition. The encoder simultaneously processes two input images (a healthy knee X-ray and an osteoarthritic knee X-ray), generating two hidden vectors for each image: one representing key features and the other non-key features. These pairs of vectors are first combined through element-wise addition to produce two hidden vectors, which are then passed to the decoder to reconstruct the corresponding outputs. Simultaneously, the key feature vector of each input is exchanged and added element-wise with the non-key feature vector of the other input. These resultant vectors are then decoded to generate the key-exchanged outputs. To achieve this, we introduce a hybrid loss strategy comprising several components: the classical AE reconstruction loss, the supervision loss from a discriminator, and a distance loss implemented using the Fisher Linear Discriminant Analysis (LDA) algorithm fisher1fisher2. The details of this hybrid loss strategy are discussed further in Section \ref{['hybrid_loss']}.
• Figure 2: The structure of the classical AE network.
• Figure 3: Classification performance obtained using SVM-RBF with different values of $\lambda_1$ and $\lambda_2$ for key feature vector $h^K$.
• Figure 4: Highlighted illustration of the original inputs and key-exchanged outputs. The green and red colours represent the absence and presence of symptoms for osteophytes and JSN, respectively. The circles and arrows represent the possible positions of the osteophytes and JSN modification, respectively.
• Figure 5: Convergence (a) and performance (b) curves obtained using various sample sizes $N$ for 500 epochs.