Built the entire 3D medical imaging vertical from scratch using MONAI — unlocking a new product category for Datature. Covers end-to-end training, data, and inference pipelines for volumetric models, with full clinical workflow compatibility via DICOM and RTSTRUCT support.
Highlights
▸End-to-end pipelines for volumetric models: SwinUNETR, SegResNet, nnUNet, SegMamba
▸NIfTI data loading, 3D augmentations, class balancing, and affine-aware evaluation
▸NIfTI annotation import/export: class-indexed and one-hot mask exports, 3D TFRecord, aggregation statistics
▸DICOM input and RTSTRUCT output for clinical radiotherapy workflow compatibility
▸Defined 3D AEMD schemas and bitmask chunk encoding in the shared core type library
▸Optimised inference with multiprocessing for mask serialisation and a bounded worker pool