Stony Brook, NY, December 18, 2025 — When a patient with cancer comes in for radiation therapy, one of the most time-consuming steps happens long before any beam is turned on. Clinicians must decide exactly how much radiation each part of the body should receive: enough to destroy the tumor, but not so much that it harms nearby healthy organs.
“This planning process can take hours and relies heavily on the experience of the individual planner,” said Zhaozheng Yin, SUNY Empire Innovation associate professor in the Department of Biomedical Informatics. “We wanted to see whether AI could provide a reference plan that could help physicians design safe, high-quality plans more efficiently.” Yin is the senior author of a new study that introduces RANDose, an AI model that predicts how radiation dose should be distributed throughout a patient’s body.
The work, developed with Ph.D. student G. Jignesh Chowdary, Dr. Tiezhi Zhang, and Dr. Xin Qian from the Department of Radiation Oncology at Stony Brook Medicine, was presented at the International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI) 2025.
From CT scans to 3D dose maps
The procedure begins with patients undergoing a CT scan. Radiation oncologists then outline the tumor and the surrounding critical organs — such as the spinal cord, salivary glands, or brainstem in head-and-neck cancers. Medical physicists use this information to run computer simulations and iteratively adjust parameters until they are satisfied with a 3D “dose map” that tells the treatment machine how much energy to deposit at each point.
The process is precise, but it is also labor-intensive and variable. Two planners may arrive at slightly different dose distributions for the same patient, especially in complex cases with many organs at risk. RANDose is designed to sit inside this workflow as a decision-support tool.
Given a patient’s CT images and outlines of the tumor and critical organs, RANDose predicts the full 3D dose distribution that a high-quality clinical plan would deliver. Clinicians can then compare their own plan against this AI-generated reference and adjust as needed. “The goal is not to replace the doctor,” Yin emphasized in the interview. “It’s to give them another reliable point of comparison.”
Teaching AI to pay attention to the right regions
What sets RANDose apart from earlier AI dose-prediction models is its focus on different parts of the anatomy. Traditional deep-learning approaches often treat every voxel — the 3D equivalent of a pixel — more or less the same during training. But in radiation therapy, not all regions are equal: the tumor, the nearby healthy tissue, and the organs at risk, all have different clinical priorities.
RANDose uses a series of attention mechanisms that explicitly encode this knowledge. One module helps the model learn how the shapes and positions of organs relate to one another in 3D space. Another focuses on how the tumor overlaps or approaches sensitive structures, which strongly influences how clinicians shape the radiation beams. Another function nudges the model to prioritize accuracy in clinically vital areas (such as the edges of the tumor) or the maximum dose to delicate organs.
This way, RANDose is able to process information at multiple spatial scales. In simpler terms, Yin said, “We’re teaching the model to look where a radiation oncologist would look when deciding what is safe and effective.”
Tests on complex head-and-neck cases
To evaluate RANDose, the team trained and tested the model on CT scans and expert-designed radiation plans for patients with head-and-neck cancers — a particularly challenging site because of the number of small, radiosensitive organs packed into a tight space.
Compared with several strong baseline models, RANDose produced dose predictions that matched more closely with the clinical plans, especially around organs at risk. The model reduced errors in the plan’s quality and better respected common clinical constraints, like limiting the maximum dose to the spinal cord. As a result, AI’s suggestions were more faithful to what experienced planners would do, while offering that guidance in seconds rather than hours.
From algorithm to clinic
For now, RANDose has been tested only retrospectively, using existing datasets. The next step, Yin said, is to integrate the model into a prototype planning system and study how it performs on new patients in collaboration with Stony Brook Medicine. That includes ensuring the tool is robust across different scanners and treatment machines, such as the CT simulators and linear accelerators used in the hospital.
Yin also sees opportunities to extend the approach beyond head-and-neck cancers to other treatment sites, and to eventually support not just dose prediction but more automated plan generation. Any such model, he stressed, would always operate under the supervision of physicians and physicists.
“Radiation therapy is ultimately about people’s lives. If we can help our clinical colleagues plan safer, faster, and more consistent treatments, we’ve found an ethical, practical way for AI to fight cancer.”