Back to research
Medical imaging AI / Quality control

WriVision: AI quality control for wrist X-rays.

This project asked a practical question: can an AI model help technologists catch wrist X-ray labeling and acquisition problems before the patient leaves? I built the page around the actual research arc: the clinical problem, the DenseNet121 multi-task model, and the places where the model worked or still struggled.

Core study facts
8,718
original wrist radiographs
6,823
filtered training set images
4
simultaneous output heads
97.23%
projection F1 score
The model was built for four checks: projection/view, laterality marker, cast presence, and hardware presence. A disagreement could trigger a correction while the patient was still in the department.
Clinical workflow

The point was quality control before the patient leaves.

Most MSK radiographs are not reviewed by a radiologist until after acquisition. That delay matters: if a wrist is mislabeled, the patient may already be gone. This system reframes the model as a near-real-time technologist assistant rather than a diagnostic black box.

Step 02

Classify

One multi-task DenseNet121 model reads the image and predicts projection, laterality, cast, and hardware at the same time.

Interactive architecture

The poster architecture, rebuilt as a cleaner full-width model map.

The key design choice was not just using a CNN. It was using a multi-task DenseNet121 backbone so the same visual representation could support four quality-control decisions. Click a layer in the diagram or selector to inspect what it contributed.

DenseNet121 flow
DenseNet121 flowClassificationProjectionLateralityCastHardwareInput imageInput image224 x 224 radiographConvolutionConvolutionPoolingPoolingDense blockDense blockx6 layersTransitionTransitionDense blockDense blockx12 layersTransitionTransitionDense blockDense blockx24 layersTransitionTransitionDense blockDense blockx16 layersFinal feature mapFinal feature mapFour output headsFour output headsprojection / laterality / cast / hardware
Active layer

Dense blocks

DenseNet121 reused features across layers, which helped one model learn related tasks instead of training four isolated classifiers.

Model performance

Four outputs, one quality-control decision.

The model only becomes useful when the outputs are interpreted together. Projection, laterality, cast, and hardware each answer a different question about whether the image is safe to send forward.

Selected task

Laterality

F1 score82.52%
PR AUC0.958
Classes

Left marker, right marker, no marker

Laterality was the clinically important stress test: the model needed to catch whether the L/R marker matched what was entered into the imaging workflow.

The hardest cases were missing or incorrectly captured markers. The no-marker precision-recall AUC dropped to 0.597, which made this the main limitation.
Training and validation

The model was tuned for generalization, not a clean-looking demo.

The dataset started with 8,718 wrist radiographs and was filtered to 6,823 de-identified images from 2,591 patients after excluding low-quantity views and enforcing label consistency.

Images were split into training/validation and test subsets with stratification based on projection. The final training setup used mini-batches of 64, Adam optimization, weighted categorical cross-entropy, learning-rate decay, and early stopping after validation loss plateaued.

Five-fold cross-validation checked that the model was not simply memorizing the training distribution. That was especially important because the class balance varied across projection, marker, cast, and hardware labels.

Five-fold rotation
A
B
C
D
E
E
A
B
C
D
D
E
A
B
C
C
D
E
A
B
B
C
D
E
A
Held-out validation fold
What I learned

The hard part was not whether the model could classify. It was whether it could be trusted in the workflow.

Marker dependence

Laterality was the most fragile output because missing or poorly captured markers were harder for the model than ordinary left/right marker reads. Future work should force the model to learn anatomy by training with marker-cropped images.

Saliency as a safety layer

Saliency maps helped show which regions influenced the prediction, especially for hardware and cast. That makes the model easier to audit before it is integrated into clinical software.

Next version

The natural extension is moving upstream: apply a similar model to a live patient-hand view before radiation is administered, then expand from wrist radiographs into other anatomy such as ankle studies where projection and laterality checks matter in a similar way.