April – August 2026 (4 months)
Research internship — data-efficient deep learning for defect detection in industrial CT
University of Bradford — Smart Manufacturing lab · Bradford, United Kingdom
A research project carried out end to end within the Faculty of Engineering & Digital Technologies, supervised by Prof. Ciprian Daniel Neagu (AIRE artificial intelligence research group). The starting question: how much real industrial data does it actually take to segment internal defects in X-ray CT automatically, and to what extent can medical pre-training or purpose-built surrogate data reduce that need?
- Designed a pipeline that generates annotated data automatically: parametric test-part generation in Python (trimesh), 3D printing at 100 % infill, CT acquisition, then ground truth derived from the scan by intensity thresholding — 90 annotated parts with no manual labelling and no CAD-to-scan registration.
- Trained the Swin-UNETR architecture (PyTorch / MONAI) for 3D volumetric segmentation: corrected an extreme class imbalance, froze then unfroze the encoder, ran sliding-window inference over volumes of several billion voxels. Dice raised from 0.33 to 0.53 on the medical calibration set.
- Ran a controlled comparison of three initialisation strategies (random, self-supervised, self-supervised plus medical fine-tuning) and plotted a data-efficiency curve from 5 to 63 parts, with multiple seeds and paired Wilcoxon tests.
- Headline result, validated on nine real industrial parts under leave-one-out cross-validation: fine-tuning on 7 real parts multiplies Dice by 1.8 to 2.1 compared with training on nine times more surrogate data.
- Wrote a 13-page research paper in English — “Data-Efficient Industrial CT Defect Segmentation via Self-Supervised and Cross-Domain Pretraining”, currently under review — and presented results weekly to my supervisors, in English.