An AI that charts a panoramic X-ray the way a dentist does: it numbers every tooth (FDI) and flags caries, deep caries, periapical lesions and impacted teeth. It then checks whether its heat-maps are anatomically plausible, and compares the AI with a dentist in a blinded study. Built on the DENTEX benchmark (MICCAI 2023).
Association Rule Mining on 10,758 EHR records from Karolinska University Hospital to learn interpretable rules for when CT or ultrasound is needed. Flagged 1,300 potentially avoidable scans — about 1,800 staff hours and 3,800 mSv of radiation.
Custom CNN vs. transfer-learned ResNet18 and MobileNetV2 for tumor vs. healthy MRI scans. ResNet18 reaches 99% validation accuracy, with Grad-CAM heat-maps to explain each prediction.
11 ML models predicting missed appointments, audited across 8 patient attributes with 4 fairness metrics. Reveals strong bias by age group and SMS-reminder status.
Decision support concept for diagnosing and managing urinary tract infections, combining EHR, lab and guideline data with HL7 FHIR and SNOMED CT. In development.