Automated Knee Osteoarthritis Grading via Compound-Scaled EfficientNetV2B0 with Attention-Based Feature Localization
Year: 2026
Type: conference-paper
Abstract: Assessment of knee osteoarthritis (KOA) by automated grading using the Kellgren Lawrence scale (KL) is crucial to enhance diagnostic consistency and the capacity to make an earlier therapeutic decision but manual analysis of radiology and computed tomography data is vulnerable to interobserver error and lack of sensitivity to subtle anatomic changes. State-of-the-art deep learning methods show promise but can be hampered by issues of class imbalance, inadequate localization of disease-relevant information, and insufficient extrapolation to diverse clinical samples. This paper presents a deep learning framework with attention enhancements that combine an EfficientNetV2B0 backbone and a person-designed attentionbased convolutional network to perform automated KL grading. The hypothesized technique exploits compound-scaled feature extraction using ImageNet pretraining as a way to enhance the efficiency of the parameters; the inverse-frequency class weighting and decision threshold calibration are used to counteract the effects of imbalance on the bias. To increase the stability of training, Adam along with AMSGrad is used to optimize. Weighing of the F1 score, the precision and the recall, as well as area under the receiver operating characteristic curve (AUC) are used to determine model performance. The outcomes are to achieve a higher classification strength, a higher sensitivity of the minority classes, and better interpretability of the result with the help of attention-directed feature representation. The suggested structure combats major weaknesses of previous methods and shows the capacity to be scaled to clinically implementable AI-supported musculoskeletal imaging procedures.
Authors: Pranjali P. Deshmukh, Pramod U. Chavan
Language: English
Topic: Medical Imaging and Analysis
Subfield: Biomedical Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7204787320