KneeGrade: knee osteoarthritis grading

Grade a knee X-ray in seconds

Two custom neural networks estimate the Kellgren–Lawrence grade, from 0 (healthy) to 4 (severe), and show which parts of the joint drove the result.

Everything runs inside your browser: X-rays never leave your device, and every analysis is saved to this browser's history.

Drop an X-ray here or click to choose a PNG or JPG file

Recent analyses

Open history

Grades in your history

About the model

Tested once on 1,656 held-out X-rays: balanced accuracy 0.644 (95% CI 0.621 to 0.666) and quadratic weighted kappa 0.753. Severe osteoarthritis is recognized in 96% of cases; separating grades 0, 1 and 2 remains the hardest part.

See model insights

X-ray

Drop an X-ray here Front view, cropped to the knee joint. PNG or JPG, up to 10 MB.

Generate a heatmap of where the model looked

No X-ray at hand? Try one of these

Shown unmodified, with their dataset labels. Source: Chen (2018), Mendeley Data, CC BY 4.0.

Search Grade From To Sort

All figures come from a single evaluation on 1,656 held-out test X-rays, run after the pipeline was frozen. Intervals are 95% bootstrap confidence intervals.

Performance by grade

Confusion matrix

Rows: true grade. Columns: predicted.

Training curves

Validation data, both ensemble members

Validation balanced accuracy

Loss, augmented baseline

Accuracy, augmented baseline

How the final pipeline was chosen

Validation set

Dataset grade distribution

Can the heatmaps be trusted?

70%of heatmap signal falls on the joint band, which covers 31% of the image

0.15 vs 0.42deletion-test area: erasing highlighted regions collapses confidence far faster than erasing random ones

−0.03similarity to an untrained network's heatmaps, so the maps reflect what the model learned

Grad-CAM heatmaps on test X-rays for each grade — Heatmaps on test X-rays: four correct and two incorrect per grade.

Deletion test chart — Deletion test on 100 test X-rays.

What KneeGrade does

KneeGrade estimates the severity of knee osteoarthritis on the Kellgren–Lawrence (KL) scale from a front-view knee X-ray, and shows a heatmap of the regions that most influenced the result. It was built as a machine-learning portfolio project: both networks were designed and trained from scratch, without pretrained weights.

Not a medical device. KneeGrade is a research and education demo. It has not been clinically validated and must not be used to diagnose or treat anyone.

The Kellgren–Lawrence scale

Dataset

8,260 knee X-rays from the Osteoarthritis Initiative, cropped to the joint at 224×224 pixels: 5,778 for training, 826 for validation and 1,656 for the final test. Grade 4 makes up only 3% of the training data, a 13-to-1 imbalance against grade 0, which is why the model is judged on balanced accuracy.

Chen, Pingjun (2018). Knee Osteoarthritis Severity Grading Dataset. Mendeley Data, V1. DOI 10.17632/56rmx5bjcr.1. Licensed under CC BY 4.0. Organized from the Osteoarthritis Initiative. Images were preprocessed and augmented for training. Neither the dataset authors nor the OAI endorse this project.

How an X-ray is graded

Standardize the image. It is converted to 224×224 grayscale. Inverted X-rays (110 in the dataset) are detected and flipped back.
Normalize contrast. Brightness is stretched using the central joint region, then CLAHE sharpens local bone edges. This removes scanner exposure differences.
Run two networks. Both predict a probability for each grade, and the two predictions are averaged. They run on your own device with TensorFlow.js, producing the same results as the original Python pipeline.
Correct for rarity. Each probability is divided by how common that grade was in training. This stops the model from defaulting to common grades, and it is why the predicted grade is occasionally not the tallest bar.
Explain. Grad-CAM highlights the regions that most influenced the predicted grade.

Model architecture

Stage	Layers	Output
Input	Grayscale X-ray	224 × 224 × 1
Block 1	2 × (3×3 convolution, batch norm, ReLU), max pooling	112 × 112 × 32
Block 2	Same pattern	56 × 56 × 64
Block 3	Same pattern	28 × 28 × 128
Block 4	Same pattern	14 × 14 × 256
Head	Global average pooling, dropout 0.3, softmax	5 grades

The first network has 1,175,845 parameters. The second uses the same design 1.5 times wider (48 to 384 filters). Both were trained with the Adam optimizer, augmentation (flips, small rotations, shifts, zoom, brightness and contrast changes), and checkpoints chosen by validation balanced accuracy.

Limitations

Trained and tested on a single source, the Osteoarthritis Initiative. Performance on other hospitals, scanners or populations is unknown.
Grade 1 ("doubtful") is not reliably separated from grades 0 and 2. On the test set, 46% of healthy knees were graded 1.
The public data has no patient identifiers, so one patient's two knees may sit in different splits, which could make results slightly optimistic.
The model grades any image it receives. Results for anything other than a front-view knee X-ray cropped to the joint are meaningless.
Heatmaps were tested for faithfulness but not checked against expert lesion annotations.

Your data

KneeGrade has no server. The models download to your browser once, and every X-ray is analyzed on your own device: images are never uploaded anywhere. Your history is saved only in this browser, so it survives refreshes but isn't shared across devices. Clearing your browser's site data, or using Clear history, removes it.

Links

Model card on Hugging Face
Source code and full write-up on GitHub