Adaptive Radiation Therapy Lab

Project description:

Delineating tumours is a bottleneck in two closely related clinical tasks: defining target volumes for adaptive radiotherapy, and measuring tumour burden on diagnostic imaging to assess response to systemic therapy. In radiotherapy, gross and clinical target volumes are still contoured largely by hand, which is slow, varies between observers and limits online plan adaptation on platforms such as MR-linacs. In oncologic imaging, response is assessed with RECIST, which reduces each patient to the diameters of a few manually selected target lesions, discards most of the available volumetric information and suffers from considerable inter-reader variability. Promptable foundation models (e.g. SAM2, MedSAM2, nnInteractive) and vision-language models now make it possible to segment structures from clicks, boxes or free-text instructions (e.g. “all liver metastases” or “left neck nodal CTV”), and large language model based agents can combine such tools with clinical context extracted from radiology and pathology reports, prior imaging and contouring guidelines. These agents will also be designed to self-assess and iteratively refine their segmentations, quantify uncertainty, and identify cases requiring expert review. This project aims to develop text-guided and agentic segmentation methods for (1) radiotherapy target volumes on CT, CBCT and MRI, including MR-guided adaptive workflows, and (2) oncologic lesions on longitudinal diagnostic CT, PET/CT and MRI. For the latter, the agent will detect and segment lesions, match them across time points and derive volumetric response categories, providing an automated alternative to manual RECIST reading while flagging uncertain cases for review. Methods will be integrated into clinical contouring software for human-in-the-loop correction and validated against expert observers and clinical RECIST reads on retrospective cohorts from LMU University Hospital. The project builds on our lab’s work on interactive AI contouring, promptable foundation models for segmentation and tracking, and our experience organising international data challenges.