Methodology
Build, train and validate an end-to-end deep-learning system that:
- Detects ankle landmarks on supine and weight-bearing X-ray images.
- Outputs the medial clear space and superior tibiotalar distance with a mean absolute error < 0.3 mm compared with expert measurements
- Runs in < 1 s on a standard workstation and shows a simple overlay of the measurement lines for visual checking.
You will start by doing a literature review of current vision-based segmentation to design a model architecture suitable for the task. Then using publicly available datasets, you will implement and using transfer learning and semi-supervised learning, pre-train your model (using datasets like the MURA musculoskeletal collection (around 40 000 limb images) and the smaller, fracture-labelled FracAtlas set (around 4000 images). Finally using Ahus ankle dataset, the model will be fine-tuned and evaluatated in its ability to supply precise landmark coordinates for the medial clear space and superior tibiotalar gap. Throughout the project, track experiments and use on-the-fly data augmentation (rotation, exposure shifts) to mimic clinical variability and learn a more generalizable model. Once the best model variant is chosen, export it to ONNX and wrap it in a simple Docker service that returns the two measurements in under a second on a standard hardware.
Expected Outcomes
At the end of this project, we aim to deliver a validated, open-source tool for automated ankle X-ray measurements, achieving sub-millimeter accuracy in under one second to improve clinical workflow. The findings will be submitted for publication. Ultimately, this tool has the potential to improve clinical workflow by providing rapid, reliable measurements, which can lead to faster clinical decision-making, reduced inter-observer variability, and a more standardized approach to patient care for ankle injuries at Akershus University Hospital and other institutions.