Confidence-Driven Deep Learning Framework for Early Detection of Knee Osteoarthritis

TL;DR

This work tackles the subjectivity in KOA KL grading by proposing a confidence-driven Siamese-GAP framework to detect early KOA (KL-0 vs KL-2). It combines multi-level feature extraction with a novel partitioned training regime and a hybrid loss that adapts to sample confidence, yielding robust performance close to expert radiologists. Key findings include competitive accuracy and F1-scores, substantial radiologist-model agreement, and improved interpretability via Grad-CAM and confidence analyses, with strong results on the OAI dataset. The approach offers a practical auxiliary tool for early KOA detection, potentially reducing clinician workload while maintaining high diagnostic reliability.

Abstract

Knee Osteoarthritis (KOA) is a prevalent musculoskeletal disorder that severely impacts mobility and quality of life, particularly among older adults. Its diagnosis often relies on subjective assessments using the Kellgren-Lawrence (KL) grading system, leading to variability in clinical evaluations. To address these challenges, we propose a confidence-driven deep learning framework for early KOA detection, focusing on distinguishing KL-0 and KL-2 stages. The Siamese-based framework integrates a novel multi-level feature extraction architecture with a hybrid loss strategy. Specifically, multi-level Global Average Pooling (GAP) layers are employed to extract features from varying network depths, ensuring comprehensive feature representation, while the hybrid loss strategy partitions training samples into high-, medium-, and low-confidence subsets. Tailored loss functions are applied to improve model robustness and effectively handle uncertainty in annotations. Experimental results on the Osteoarthritis Initiative (OAI) dataset demonstrate that the proposed framework achieves competitive accuracy, sensitivity, and specificity, comparable to those of expert radiologists. Cohen's kappa values (k > 0.85)) confirm substantial agreement, while McNemar's test (p > 0.05) indicates no statistically significant differences between the model and radiologists. Additionally, Confidence distribution analysis reveals that the model emulates radiologists' decision-making patterns. These findings highlight the potential of the proposed approach to serve as an auxiliary diagnostic tool, enhancing early KOA detection and reducing clinical workload.

Figures (8)

• Figure 1: The flowchart of the proposed approach consists of two main components: the Siamese-GAP model and the hybrid loss strategy. The overall data flow is represented by black arrows. On the left, the Siamese-GAP network includes red vertical arrows to denote GAP units, blue braces indicating concatenation operations, the $\oplus$ symbol representing element-wise addition, and a purple block for the fully connected layer followed by Softmax. On the right, the hybrid loss strategy showcases the data flow of validation and training batches with light purple and light blue arrows, respectively. Samples within each KL grade are ranked by their confidence level using $\lambda$, while the three loss functions are weighted by hyper-parameters $\alpha$, $\beta$, and $\gamma$. Additional parameters, such as $c_0$ and $c_2$, are detailed in Section \ref{['hyperparaters']}.
• Figure 2: The structure of the classical Siamese network.
• Figure 3: Lateral and medial patches are in blue and orange boxes (a). Obtained patches serve as input of the model (b).
• Figure 4: Comparison of the attention maps of the evaluated models.
• Figure 5: Comparison of confidence distributions. Subfigure (a) shows the distribution under the single CE loss function, while subfigure (b) illustrates the distribution using the hybrid loss strategy. The confidence scores are sorted in descending order.