← work

Medical imaging · experimental research

Learning to segment small structures in CT volumes

GitHub repository

A segmentation research pipeline that aligns DICOM volumes with NRRD annotations and explores how cropping, model architecture, loss functions, and ensembles affect small-structure segmentation.

The implementation includes patient-level data splitting, CT intensity windowing, localization and refinement stages, and a 3D U-Net. Experiment scripts cover foreground sampling, augmentation, loss combinations, postprocessing, and ensemble evaluation.

The experiment log describes 52 patients and reports both Dice and normalized surface-distance scores. It records 0.62 Dice for one five-fold ensemble setting and 0.64 for a separate single-split 3D U-Net experiment. These protocols are different and should be read separately. This is an experimental research system, without a claim of clinical validation.