ClinNet: Evidential Ordinal Regression with Bilateral Asymmetry and Prototype Memory for Knee Osteoarthritis Grading Journal: Medical Image Analysis
Xiaoyang Li, Runni Zhou · 2026
Read ClinNet: Evidential Ordinal Regression with Bilateral Asymmetry and Prototype Memory for Knee Osteoarthritis Grading Journal: Medical Image Analysis by Xiaoyang Li, Runni Zhou (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-17315.txt.
Introduction
Xiaoyang Li and Runni Zhou present ClinNet, a tool to grade knee osteoarthritis from X-ray images. Grading is hard for three reasons. Grades differ by small signs. Human labels carry real doubt. And the grades form an order, from mild to severe. Most past tools treat the task as plain class choice. That misses the order and the doubt.
ClinNet uses evidential ordinal regression to keep both. It adds a check on left-right asymmetry in the knees. It also keeps a prototype memory of past cases. The mix aims for grades a clinician can trust. Better grading guides care for a joint disease that affects many older adults.
Catalog record
| CREATOR | Xiaoyang Li, Runni Zhou |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 46 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2601-17315
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
ClinNet: Evidential Ordinal Regression with Bilateral Asymmetry and Prototype Memory for Knee Osteoarthritis Grading Journal: Medical Image Analysis Xiaoyang Li, Runni Zhou Abstract Knee osteoarthritis (KOA) grading based on radiographic images is a critical yet challenging task due to subtle inter-grade differences, annotation uncertainty, and the inherently ordinal nature of disease progression. [ 25 , 7 , 26 , 17 ] Conventional deep learning approaches typically formulate this problem as deterministic multi-class classification, ignoring both the continuous progression of degeneration and the uncertainty in expert annotations. [ 50 , 3 , 16 , 27 ] In this work, we propose ClinNet, a novel trustworthy framework that addresses KOA grading as an evidential ordinal regression problem. [ 2 , 44 , 5 ] The proposed method integrates three key components: (1) a Bilateral Asymmetry Encoder (BAE) that explicitly models medial–lateral structural discrepancies; [ 1 , 54 , 23 ] (2) a Diagnostic Memory Bank that maintains class-wise prototypes to stabilize feature representations; [ 46 , 55 , 19 ] and (3) an Evidential Ordinal Head based on the Normal-Inverse-Gamma (NIG) distribution to jointly estimate continuous KL grades and epistemic uncertainty. [ 2 , 27 ] Extensive experiments demonstrate that ClinNet achieves a Quadratic Weighted Kappa of 0.892 and Accuracy of 0.768, statistically outperforming state-of-the-art baselines ( p < 0.001 p<0.001 ). [ 6 , 13 ] Crucially, we demonstrate that the model’s uncertainty estimates successfully flag out-of-distribution samples and potential misdiagnoses, paving the way for safe clinical deployment. [ 22 , 41 , 15 , 12 ] 1 Introduction 1.1 Clinical Background & Motivation Knee osteoarthritis (KOA) is a chronic and progressive degenerative musculoskeletal disease and a leading cause of pain, mobility impairment, and functional disability among older adults worldwide. [ 25 , 7 ] With the rapid aging of the global population, the prevalence of KOA continues to increase, placing a substantial burden on both patients and healthcare systems. [ 7 ] In routine clinical practice, radiography remains the most used imaging modality for KOA assessment due to its low cost and wide availability. [ 25 , 1 ] Among existing diagnostic criteria, the Kellgren–Lawrence (KL) grading system has been widely adopted as the gold standard for evaluating radiographic disease severity. [ 26 , 1 ] Despite its widespread use, KL grading is inherently subjective, as it relies on visual assessment of multiple radiographic features, including joint space narrowing, osteophyte formation, subchondral sclerosis, and bony deformities. [ 26 , 1 ] These features typically emerge gradually and heterogeneously, making precise grade assignment challenging. [ 25 ] Prior studies have consistently reported notable inter-observer and intra-observer variability in KL grading, particularly in the early stages of the disease, where radiographic changes are subtle and poorly defined. [ 30 , 53 ] The ambiguity is most evident when differentiating KL grade 1 (doubtful KOA) from KL grade 2 (minimal KOA). [ 26 , 53 ] Conceptually, KL-1 is characterized by possible osteophytic lipping, whereas KL-2 requires definite osteophytes with possible joint space narrowing. [ 26 , 1 ] In practice, however, the distinction between “possible” and “definite” osteophytes lacks a clear objective boundary. [ 1 ] Early osteophytes are often small, low-contrast, and sensitive to imaging conditions such as projection angle and exposure, while mild joint space changes may be confounded by normal anatomical variability. [ 1 , 53 ] Even experienced musculoskeletal radiologists frequently encounter difficulty in consistently distinguishing between these adjacent grades. [ 53 ] This uncertainty has important clinical implications, as the transition from KL-1 to KL-2 often marks the threshold for a radiographic diagnosis of KOA and may influence subsequent monitoring and intervention strategies. [ 25 ] Radiographic assessment of KOA further relies heavily on intra-articular anatomical comparison within a single knee. [ 1 ] In particular, the relative difference between the medial and lateral compartments plays a critical role in clinical interpretation. [ 1 ] Due to asymmetric biomechanical loading, KOA most commonly affects the medial compartment first, leading to earlier joint space narrowing and osteophyte formation compared with the lateral compartment. [ 25 , 1 ] Consequently, radiologists frequently assess compartmental asymmetry to determine whether observed changes reflect pathological degeneration rather than normal variation, placing greater emphasis on relative structural differences than on absolute measurements alone. [ 1 ] However, this clinically grounded reasoning is often insufficiently captured by existing automated KOA grading methods. [ 49 , 33 ] Many deep learning approaches rely on cropped images of a single knee and treat the task as a generic image classification problem. [ 50 , 3 , 49 ] Such formulations tend to focus on localized appearance or global texture features, while failing to explicitly model structural relationships between the medial and lateral compartments. [ 33 ] This limitation is particularly detrimental for early-stage KOA, where absolute radiographic changes are minimal and subtle compartmental asymmetries provide crucial diagnostic cues. [ 1 , 53 ] Moreover, the KL grading system represents an ordinal and progressive disease continuum, rather than a set of independent categorical labels. [ 26 , 17 ] Adjacent grades differ incrementally, yet many existing approaches formulate KOA grading as a standard multi-class classification task, disregarding the inherent ordinal structure. [ 50 ] This mismatch can lead to predictions that are inconsistent with clinical understanding of disease progression and limits the reliability of automated systems in real-world applications. [ 39 , 5 ] Motivated by these clinical considerations, this study aims to develop an automated KOA assessment framework that better aligns with radiological practice by explicitly modeling the ordinal nature of KL grades and incorporating anatomically informed representations of medial–lateral compartmental structure. [ 17 , 1 ] Such an approach seeks to improve early-stage discrimination, robustness, and clinical interpretability, thereby enhancing the translational potential of automated KOA grading systems. [ 27 , 41 ] 1.2 Related Work 1.2.1 Traditional Radiographic Feature–Based KOA Assessment Early efforts in automated knee osteoarthritis (KOA) assessment were primarily based on handcrafted radiographic features. [ 1 ] Commonly used descriptors included joint space width (JSW), osteophyte size and shape, subchondral bone density, edge irregularity, and texture-based statistics derived from predefined regions of interest. [ 1 ] These features were typically combined with conventional machine learning models such as support vector machines, k-nearest neighbors, or random forests to perform KOA detection or severity grading. [ 42 ] While these approaches enabled quantitative analysis of radiographic characteristics, their performance was heavily dependent on the robustness of feature extraction pipelines. [ 25 ] Variations in imaging acquisition, patient positioning, and anatomical morphology often led to unstable measurements. [ 1 ] Moreover, handcrafted features generally capture isolated aspects of joint degeneration and fail to represent the complex, multifactorial progression of KOA. [ 25 , 7 ] As a result, such methods demonstrated limited generalization ability and were gradually supplanted by data-driven deep learning approaches. [ 33 ] 1.2.2 CNN-Based Automated KOA Grading With the rapid advancement of deep learning, convolutional neural networks (CNNs) have become the dominant paradigm for automated KOA grading from knee radiographs. [ 33 , 50 , 3 , 49 ] Numerous studies have adopted backbone architectures such as ResNet, DenseNet, EfficientNet, and ConvNeXt to learn discriminative features in an end-to-end fashion. [