Start of funding 01.07.2026

Thinking Beyond Recognition: Reliable and Generalizable Radiology Report Generation with Test-Time Thinking

Prof. Dr. Julia Schnabel
Technische Universität München
Computational Imaging and AI in Medicine (CompAI)

Prof. Dr. Akshay Chaudhari
Stanford University
Stanford Center for Artificial Intelligence in Medicine and Imaging (AIMI)



Recent advances in AI highlight that Vision-Language Models (VLMs) can improve decision quality by reasoning before responding at inference time, refining intermediate decisions, and verifying outputs before producing a final answer. This growing evidence suggests that test-time thinking by models during inference improves output quality without retraining, making it especially promising for radiology report generation, where these models must align findings with image evidence, remain robust to unseen abnormalities, and produce clinically coherent reports. This project, between the CompAI lab at TU Munich and the MIMI lab at Stanford University, aims to develop a test-time thinking method for radiology report generation that treats inference as structured report construction, refinement, and verification, rather than single-pass generation. By leveraging test-time thinking within fixed and limited compute, this project aims to make radiology report generation more reliable and clinically dependable, without the cost of excessive retraining or scaling the model.