Research software, applications and reproducible pipelines — from medical imaging to biosignals and open-source tools
Journal project building an uncertainty-aware deep learning classifier for central-canal stenosis grading from lumbar sagittal MRI. Combined Monte-Carlo dropout and test-time augmentation to flag low-confidence predictions for clinical review. Published in JOR Spine, 2026.
Data-driven study identifying which chronic low-back-pain patients benefit from lumbar steroid injections. Developed nested cross-validated ML models with SHAP interpretability on the TREXI cohort (212 patients), reaching 0.865 average precision and surfacing pain self-efficacy as the strongest predictor. Published in Scientific Reports (Nature).
Diffusion-based deep-learning framework (ETH Zürich BMDS Lab) for automated multiclass segmentation of lumbar spine MRI in low-back-pain patients. Segments vertebral bodies, discs and the spinal canal from T1- and T2-weighted scans, with a presegmentation strategy that cuts diffusion timesteps from 1000 to ~30 while matching nnU-Net accuracy and producing uncertainty heatmaps. Published at MIDL 2025.
Open-source contribution to the ORMIR community: a specification and Python package that standardises how musculoskeletal imaging data is organised, curated and shared — a DICOM→ORMIR-MIDS converter plus a Python API to find, load and interrogate data in a BIDS-style layout for reproducible research.
Small Android demo application which classifies singing vs speaking vs silence over the last 2 seconds based on microphone sensor data.
Web app that brings a computer-science department exhibition online — a panorama-based 3D viewer, built in an AMOS agile/Scrum team.
Deep-learning pipeline to decode reach-and-grasp motor intention from EEG brain signals for brain-computer interface research at Bit Brain.
Two chained CNNs for automatic mitral valve-annulus landmark tracking on 2- and 4-chamber CINE CMR, with an interactive dashboard for cluster analysis and anomaly detection. Co-invention basis for patent US12239477B2 (presented at ISMRM 2021).
Illustrative example of a Reinforcement Learning algorithm for optimal storing and pick-up routing of objects in a warehouse.
M.Sc. Thesis carried out in collaboration with Siemens Healthineers. Research state-of-the-art literature on cardiac MRI slice tracking and medical landmark detection. Develop a robust novel valve-tracking 3D-CNN algorithm based on heatmap regression.
Deep-learning segmentation combining bSSFP, T2 and LGE cardiac MRI to delineate infarcted and edema myocardial regions for myocardial-infarction diagnosis.
Multinomial Naive Bayes text classifier that categorizes financial transactions into seven types.
New adaptive sampling strategy based on a motion-context measure for egocentric video was developed. The new method enables to save energy while keeping a promising performance rate.
MATLAB GUI for ECG annotation plus an SVM classifier that retrospectively detects arrhythmia during resuscitation episodes from five discriminative ECG features. BSc thesis.