Source-linked AI summary
Automatic Knee Osteoarthritis Diagnosis from Plain Radiographs: A Deep Learning-Based Approach
Aleksei Tiulpin, Jérôme Thevenot, Esa Rahtu, Petri Lehenkari, Simo Saarakkala
TL;DR
Early knee OA is difficult to diagnose from plain radiographs, which are insensitive to early changes and require interpretation of two-dimensional images. This study develops a neural-network approach that learns disease-relevant features and provides attention maps, achieving strong OA grading performance across datasets.
Problem
Plain radiography is widely available but insensitive to early OA changes, while early diagnosis matters because progression may still be slowed or its future disability reduced.
Method
The study develops an efficient neural-network architecture that learns relevant OA features and uses GradCAM-based class-discriminating attention maps for supplementary diagnostic information.
Results
66.71% average multi-class accuracy, 0.93 radiographical OA diagnosis AUC, 0.83 quadratic weighted Kappa, and 0.48 MSE were achieved on the reported evaluation.
Takeaways & Limitations
Attention maps provide supplementary information about localized radiographical findings and may help build clinical trust, while the released dataset supports research comparison across automatic OA grading methods.
Takeaways & Limitations
The study used an OAI validation set and reduced image resolution to 8 bits, potentially losing fine-grained image information; broader generalisability remains for future study.
Abstract
from arXiv · showhide
Knee osteoarthritis (OA) is the most common musculoskeletal disorder. OA diagnosis is currently conducted by assessing symptoms and evaluating plain radiographs, but this process suffers from subjectivity. In this study, we present a new transparent computer-aided diagnosis method based on the Deep Siamese Convolutional Neural Network to automatically score knee OA severity according to the Kellgren-Lawrence grading scale. We trained our method using the data solely from the Multicenter Osteoarthritis Study and validated it on randomly selected 3,000 subjects (5,960 knees) from Osteoarthritis Initiative dataset. Our method yielded a quadratic Kappa coefficient of 0.83 and average multiclass accuracy of 66.71\% compared to the annotations given by a committee of clinical experts. Here, we also report a radiological OA diagnosis area under the ROC curve of 0.93. We also present attention maps -- given as a class probability distribution -- highlighting the radiological features affecting the network decision. This information makes the decision process transparent for the practitioner, which builds better trust toward automatic methods. We believe that our model is useful for clinical decision making and for OA research; therefore, we openly release our training codes and the data set created in this study.
1 Introduction
Knee OA is common and costly, but early diagnosis from plain radiographs is difficult because radiography misses subtle changes and KL grading is subjective. The study proposes a transparent CADx approach using disease-relevant features, attention maps, Siamese networks, cross-dataset evaluation, and a released benchmark dataset.
- Motivation: OA is the world’s most common musculoskeletal disorder, with substantial disability and treatment costs.Hip and knee OA are described as the eleventh highest global disability factor, with estimated treatment costs reaching 19,000 €/year per patient.
- Motivation: Plain radiography is safe, inexpensive, and widely available, but its two-dimensional projections and indirect cartilage assessment make early OA difficult to detect.Radiographs infer cartilage degeneration through joint-space narrowing and bony changes such as osteophytes and subchondral sclerosis.
- Motivation: KL grading is semi-quantitative and ambiguous, producing substantial reader disagreement and motivating objective CADx support for clinicians.Reported reader agreements include quadratic Kappa values of 0.566, 0.667, and 0.678.
- Motivation: Deep learning can achieve strong CADx performance, but its black-box decisions and need for large datasets limit clinical interpretability and validation.The introduction emphasizes that transparency is crucial before clinical use and that models should be tested on data distinct from training data.
- Approach: The proposed method combines a disease-specific CNN, KL-grade probability distributions, and GradCAM attention maps to support transparent OA diagnosis from radiographs.The model highlights radiological features contributing to its decisions while representing grading uncertainty as a probability distribution.
- Contributions: The study introduces a Siamese architecture for symmetric medical images, releases a standardized knee X-ray dataset, and evaluates transferability on data not used for training.The stated novelties include reduced learnable parameters, a public benchmark dataset, and cross-dataset evaluation.
2 Materials and Methods
The study uses separate MOST data for model development and OAI baseline data for testing, with knee-region preprocessing and a symmetry-based Deep Siamese CNN. The system combines three models to predict five-grade KL probability distributions while focusing on lateral and medial joint features.
- Datasets: MOST supplied model-development data, while the OAI baseline supplied testing data from different subjects.The cohorts contain KL-graded knee radiographs from adults in the stated age ranges; OAI was not used for training.
- Pre-processing: The pipeline localised knee joints, cropped lateral and medial patches, and horizontally flipped left-knee or medial images to standardise side appearance.Patch extraction used knee-region coordinates and validation-selected preprocessing parameters.
- Network architecture: The Deep Siamese CNN used two shared-weight branches whose outputs were concatenated for KL-grade prediction.The architecture combined convolution, batch normalisation, ReLU, max-pooling, global average pooling, concatenation, and a final prediction layer.
- Network architecture: The model constrained attention to the knee’s lateral and medial regions rather than selecting image locations adaptively.This design used known knee structure to learn local features relevant to fine-grained OA grading.
- Inference: Three models trained with different random seeds were selected using validation performance and their predictions were fused before softmax normalisation.The ensemble was intended to imitate multiple-reader grading while producing a KL-grade probability distribution.
- Design rationale: The fine-grained KL-grading task and limited training size created an overfitting risk to background noise or image artefacts.The study therefore emphasised learning disease-relevant local radiographic features.
3 Experiments and Results
The method was evaluated on held-out OAI knees after model selection and ensemble training, achieving strong agreement and diagnostic performance while producing interpretable attention and confidence outputs.
- 66.71% average multi-class accuracy and 0.93 radiographical OA diagnosis AUC were achieved on the OAI test set.
- 0.83 quadratic Kappa agreement was obtained against expert annotations, while classification MSE was 0.48.The Kappa metric weights larger grading disagreements more heavily than smaller ones.
- The fine-tuned ResNet-34 baseline had similar overall test performance but reached 67.49% average accuracy and 0.51 MSE.Its Kappa agreement was also 0.83, and the proposed method performed better for KL-2 cases, 52% versus 48%.
- The proposed model’s attention maps focused on local radiological findings such as osteophytes, unlike the baseline’s attention to joint centres, bone texture, or background regions.The architecture’s domain-knowledge constraints targeted osteophytes, bone deformity, and joint-space narrowing used in KL grading.
- The network also outputs a probability distribution over KL grades, providing supplementary information about prediction confidence and diagnostic uncertainty.The distribution allows practitioners to see when predictions are less confident rather than receiving only a finite grade.
4 Discussion
The discussion emphasizes cross-dataset robustness, clinically relevant interpretability, and practical utility, while identifying validation, resolution, and generalisability limitations.
- Training on MOST and testing on OAI demonstrated transfer of relevant OA features across datasets and acquisition settings.The study describes this design as supporting robustness toward different artefacts and data acquisition settings.
- The method uses normalised regions of interest, constrained attention zones, varied beam angles, augmentation, and three-network ensembling to improve robustness.The ensemble used different random seeds to reduce variance in model decisions.
- The validation design reused OAI data for validation, and future work is needed to assess generalisability across multiple datasets with larger samples.The discussion identifies this as a study limitation rather than a definitive failure of the method.
- Reducing radiographs to 8-bit resolution may have removed fine-grained information, while baseline attention maps had lower spatial resolution than the proposed model’s maps.The authors suggest original-resolution images with reasonable filtering could further improve results.
- The attention maps and KL-grade probability distributions provide clinicians with localised findings and supplementary information for interpreting uncertain severity estimates.The authors suggest these outputs may help inexperienced practitioners and support their training.
- The approach is presented as useful for faster diagnosis, reducing routine healthcare costs, monitoring other knee conditions, and analysing large OA cohorts.The authors also release bounding-box datasets, pretrained models, and training codes as a benchmark for automatic grading methods.
Additional information
The authors declare no competing financial interests.
- The authors declare no competing financial interests.
5 Supplementary Information
The supplementary section details training, model comparison, and qualitative attention-map analyses for the proposed knee OA grading approach. It addresses class imbalance, evaluates competing models, and examines how anatomical constraints affect network attention.
- Training implementation: Oversampling, data augmentation, and batch-selection strategies were used to address class imbalance and help prevent overfitting.Training categories KL 0–4 were sampled with repetitions, while augmentation varied brightness, contrast, rotation, gamma, and jitter.
- Model selection: The comparison included a re-implemented classification-regression network, fine-tuned ResNet-34, and multiple configurations of the proposed models.The re-implemented network used classification and regression outputs optimized with averaged mean squared error and cross-entropy.
- Model selection: The proposed model with N = 64 filters and the fine-tuned ResNet-34 were selected as the two best models based on validation Kappa.Table 2 identifies these models as the two best according to its Kappa column.
- Attention-map analysis: Attention maps were compared for clinically relevant KL-2 cases and moderate OA cases to assess model focus on radiological features.The analyses included Figures 6 and 7 and compared the fine-tuned ResNet-34 with the proposed model.
- Attention-map analysis: Constraining network attention with prior anatomical knowledge was associated with a benefit in the attention-map examples.For KL-2, the pre-trained model was described as less certain than the proposed approach.