Interactive 3D segmentation toolset
AstraZeneca · Medical Imaging & Segmentation
Problem
Manually annotating tumor regions on 3D CT volumes is the bottleneck for training and evaluating downstream segmentation models. Domain experts spend substantial time per volume, and inter-annotator variability adds noise to both ground truth and evaluation. We needed a tool that gave annotators strong AI-assisted starting points while keeping them firmly in control of the final segmentation.
Approach
Interactive segmentation built on a transformer-backed architecture, with two interaction modes:
- Click-based prompts — a small number of foreground/background clicks produce an initial segmentation, refinable through additional clicks.
- Text-guided prompts — natural language descriptions of anatomy or pathology condition the segmentation, useful when the annotator wants semantic disambiguation.
Wrapped in a web-based annotation UI with versioned outputs, audit trails, and integration into the team’s broader imaging data platform.
Result
Halved annotation time on representative cases. Deployed to twelve internal users across R&D, with companion paper as poster at AACR 2024:
Patwari M, Wei Y, Xu M, Zhang Z, Sidiropoulos K, Selvaraj B, Hughes G, et al. “Fast, interactive, AI-assisted 3D lung tumour segmentation.” Poster, AACR Annual Meeting 2024. Abstract: Cancer Research 84, Suppl. 6, 887.
Stack
PyTorch, transformer-based foundation segmentation backbone, MONAI for pre/postprocessing, custom React/Dash frontend for annotation workflow.