Dual-Model Approach for Knee Osteoarthritis Classification Using Custom CNN and Fine-Tuned VGG16 with Histogram Equalization

Abstract

Knee osteoarthritis (KOA) is a progressive joint disorder that often remains asymptomatic in early stages, making timely diagnosis essential. This study presents a deep learning-based framework for KOA classification using preprocessed knee X-ray images. Image enhancement techniques such as histogram equalization and adaptive histogram equalization were applied to improve local contrast. Two independent architectures were developed: a custom convolutional neural network (CNN) designed from scratch and a fine-tuned VGG16 model based on transfer learning. The custom CNN incorporated L1/L2 regularization, batch normalization, and dropout layers to improve generalization. However, initial results revealed reduced performance due to severe class imbalance in the dataset. To address this, data augmentation was employed as an oversampling strategy, using rotation, flipping, translation, shearing, and zoom transformations to balance all KOA grades. To further improve classification performance, an ensemble strategy was introduced by combining the softmax probability outputs of the custom CNN and VGG16 using weighted averaging. This fusion allows the complementary strengths of both models to jointly influence the final Kellgren–Lawrence grade prediction, improving robustness and early-stage KOA detection. Although augmentation increased data diversity, the custom CNN achieved limited accuracy, prompting the adoption of the more robust VGG16 architecture. Comparative evaluation showed that VGG16 and the proposed ensemble achieved superior classification performance, demonstrating the effectiveness of transfer learning for KOA detection. The proposed framework offers a reliable approach for assisting early KOA diagnosis and supporting clinical decision-making. Received: 27 October 2025 | Revised: 29 January 2026 | Accepted: 13 February 2026

Conflicts of Interest

The authors declare that they have no conflicts of interest to this work.

Data Availability Statement

This study utilized data from the OAI database, a public resource available at https://nda.nih.gov/oai/. The OAI is a multicenter, longitudinal study of KOA, funded by the National Institutes of Health.

Author Contribution Statement

G. Prema: Conceptualization, Methodology, Validation, Formal analysis, Investigation, Resources, Data curation, Writing – review & editing, Visualization, Supervision, Project administration. P. Rithanya: Software, Resources, Data curation, Writing – original draft, Writing – review & editing, Visualization. A. Shiva Priya: Software, Resources, Data curation, Writing – original draft, Writing – review & editing, Visualization.