The room

Projects

Medical imaging AI from my work at MED-EL, plus personal projects in MRI segmentation, computer vision and assistive technology. Everything public lives on GitHub.

Professional and research software

2023–present · Python · PyTorch · MONAI · C++

CI-Planner (formerly OTOAI): inner-ear detection, segmentation and classification

A research-use-only decision-support platform for inner-ear CT and MRI, built with the four-person team I lead. Three deep-learning models run in one workflow: detection, segmentation of the inner ear and surrounding structures, and classification of normal versus pathological anatomy. Early cropping algorithms (HU-based and clustering-based) isolate the inner ear from a full DICOM scan, and a 3D residual U-Net reached a best validation Dice of 0.88 on a small held-out set. The models run on-premises in a desktop application through ONNX Runtime.

  • 3D U-Net
  • 3D classifier
  • MONAI
  • SimpleITK
  • DICOM
  • ONNX Runtime
  • PySide6

Personal projects

2023, rebuilt 2026 · Python · TensorFlow · Keras

Brain MRI tumour segmentation with U-Net

Segments lower-grade glioma on brain MRI slices (TCGA-LGG, 3,929 slices from 110 patients). I rebuilt an earlier notebook experiment as a tested Python package with training and evaluation scripts, an inference API and a Streamlit demo. Evaluation is split by patient: on a held-out test set of 16 patients the default 1.95 M-parameter U-Net reaches Dice 0.87 over all slices and 0.70 on slices that contain a tumour, at about 10 ms per slice on a GTX 1650 Ti. The notebook's original 0.91 Dice leaked data between train and test, so the repository documents what was fixed. Research use only.

  • U-Net
  • MRI segmentation
  • TensorFlow / Keras
  • Streamlit
  • TCGA-LGG

2022 · Python · OpenCV

Smart Glasses for the visually impaired

A wearable assistant controlled entirely by voice in Arabic. It has three modes: it recognises registered people by face and says their name, names the objects in view with YOLOv3 and whether they are on the left or right, and reads Arabic signs such as pharmacy or clinic aloud.

  • YOLOv3
  • LBPH face recognition
  • Arabic OCR
  • Speech recognition
  • Text to speech

2021 · Python · Deep learning

Gaze tracker for autism research

A prototype for studying gaze in children with autism, from live camera or recorded video. I built a desktop app around the open-source DeepVOG pupil-segmentation network, with a child-info form and a calibrate-then-infer workflow that fits an eyeball model and estimates gaze direction.

  • DeepVOG
  • Keras
  • Tkinter
  • Eye tracking

Python · Computer vision

Facial expression analyzer

Finds a face, locates 68 facial landmarks and classifies the expression with an SVM, returning a score for each emotion. A small Tornado web service accepts a photo and replies with the emotion and score, with a leaderboard endpoint for an expression game.

  • dlib
  • scikit-learn
  • OpenCV
  • Tornado

See certifications and awards →