Lightweight Vision Transformer Ensemble With External Validation For Binary Knee Osteoarthritis Detection

Abstract

Knee osteoarthritis (KOA) is another degenerative joint disorder that has a profound effect on mobility and quality of life especially with aging populations. Timely detection may rely on early and accurate detection, yet the traditional method of grading radiological images by manual means is subjective, non-objective, and likely to generate inter-observer variability. In order to alleviate these constraints, this paper introduces a lightweight Vision Transformer (ViT) system to compute automated binary KOA severity stratification No/Mild vs. Moderate/Severe in relation to clinically acting on treatment thresholds. The implementation is based on a group of 3 ViT-Small models trained with different random encycles, test-time augmentation to make the models more robust, and measures performance by rigorously stratifying the patients with external validation. Pubically available knee MRI datasets were experimented on, and binary labels were based on Kellgren-Lawrence grades. The performance of the proposed model on internal testing was 87.5% and the external validation was 87.09 which showed a low performance decline of 0.41% which is much lower than the general performance degradation observed in the KOA literature. It is noteworthy that the framework achieved a high sensitivity (97.56s) in women in moderately severe cases, low violence inference latency (6.56 ms/image) and interpretable attention maps, which are used to emphasize parts of the anatomy of interest. These results: When parameter-efficient transformer architectures are used along with parameter-efficient ensemble learning, and trained with stricter validation procedures, these methods can be able to produce robust, generalizable, and clinically interpretable KOA screening tools. This paper highlights the significance of external validity and computational efficiency to apply medical AI models to the real-world healthcare environment to provide a repeatable base to develop multimodal extensions and anticipate clinical application.