Hip osteoporosis risk from CT imaging and clinical data
Lead researcher, under Dr. Champa Jayasuriya
A prior result did not survive honest re-evaluation, and a simpler model proved more trustworthy.
0.976DiceU-Net bone segmentation
The problem
Osteoporosis usually goes undiagnosed until a fracture happens. A hip CT taken for another reason already contains bone-density information; the question is whether it can be combined with routine clinical variables to flag risk earlier.
The data
109 hip CT patients — 71 with osteoporosis, 38 without — from University of Toledo Medical Center, with 15 clinical variables including age, BMI, calcium, haemoglobin and glucose.
The imaging is real patient data and never leaves the institution. Nothing derived from it is published here beyond segmentation masks and schematics.
A confound found during the audit: some images had been exported at different scan settings, which can bias any texture-based feature.
Approach
U-Net segmentation to isolate bone regions, reaching a Dice coefficient of 0.976, then radiomic feature extraction with PyRadiomics — texture, shape and intensity — fused with the clinical variables.
Six models trained under Bayesian hyperparameter optimisation: logistic regression, SVM, random forest, AdaBoost, CatBoost and a DNN.
The pipeline was inherited from a prior student and audited end to end before any new result was produced.
Results
The audit found a data-leakage bug: the model was effectively seeing test data during training.
After correction and re-validation, a previously reported headline result did not hold. A simpler, more robust model emerged as the most trustworthy of the six.
A single train/test split was replaced with repeated cross-validation, which showed the initial results were noise-driven and confirmed genuine signal only under stricter testing.
Final metrics for the corrected models belong to the thesis and are not published here. No number is quoted in their place.
Limitations and what I'd do differently
109 patients is a small cohort with an imbalanced split, 71 to 38. Repeated cross-validation gives a distribution rather than a point estimate, but it cannot manufacture statistical power that the sample size does not contain.
The scan-setting confound is flagged, not solved. Stratifying by acquisition parameters, or harmonising features across them, is the correct next step and has not been done.
Single-centre data. Nothing here demonstrates that the result transfers to another hospital’s scanners or population.
The lesson I would carry forward: audit the inherited pipeline before extending it. The leakage was upstream of everything I was asked to add.
Artifacts
One paper published; citation to follow. Patient imaging is never published in any form.