Covering Scientific & Technical AI

Researchers at UC San Diego have developed an artificial intelligence tool that significantly reduces the time required to plan brachytherapy treatments for cervical cancer, an effective but labor-intensive form of internal radiation therapy.

The system, described in a recent paper in the journal Brachytherapy, integrates a deep learning model directly into a commercial treatment planning system, allowing clinicians to generate customized treatment plans with a single click. The approach was developed using the Voyager supercomputer at the San Diego Supercomputer Center (SDSC), with support from the National Science Foundation’s NAIRR Pilot program. According to the researchers, access to shared national cyberinfrastructure was critical for training, testing, and integrating the model into a clinically usable workflow.

SDSC’s Voyager is an NSF-funded supercomputer built with advanced Intel/Habana Gaudi processors that are optimized for AI applications (Image and Caption Credit: SDSC)

“We used NAIRR Pilot allocations on SDSC’s Voyager system to develop this new tool and built it into a widely used medical software platform so it’s easy for clinicians to utilize it,” said Lance Moore, an AI researcher with the UC San Diego School of Medicine Department of Radiation Medicine and Applied Sciences, in an SDSC article. “The new tool uses AI to automate and speed up treatment planning and with just one click, the system analyzes a patient’s medical images and creates a high-quality, customized plan in less than four minutes — potentially reducing both the patient’s discomfort and the risk of human error.”

Cervical cancer affects roughly 600,000 women worldwide each year and causes an estimated 340,000 deaths, according to the World Health Organization. Brachytherapy is a standard option of care for many patients, but treatment planning is complex and time-consuming, often taking an hour or more and requiring extensive clinician expertise. The UC San Diego team sought to address these challenges by automating the most time-consuming steps of treatment planning while preserving clinical accuracy. Their system combines a deep learning dose prediction model with an optimization pipeline that converts predicted dose distributions into deliverable treatment plans. Unlike prior research workflows that require manual export and re-import of medical imaging data, the new tool operates entirely within Varian’s BrachyVision treatment planning system using an API.

At the heart of the tool is a 3D cascade U-Net model trained on more than 1,800 historical cervical brachytherapy plans from UC San Diego. The model predicts a patient-specific radiation dose distribution based on CT images, anatomical contours, and the physical layout of the brachytherapy applicator inside the body. An optimization step then determines dwell times, or how long the radioactive source remains at each position, to closely match the predicted dose.

This two-stage AI model (called cascade U-Net) predicts the radiation dose for a brachytherapy plan. It takes simple images of the tumor, nearby organs and other planning images in order to make a patient-specific prediction (Credit: Lance Moore)

In tests on 28 cervical cancer cases covering seven different applicator designs, the automated system produced treatment plans that closely tracked those used in clinical practice. On average, the AI-generated radiation maps differed from clinically delivered plans by less than 4% across the 3D treatment area, and plans were generated in about three and a half minutes. More complex cases took longer to process, but planning times remained far shorter than traditional manual workflows.

By reducing planning time and variability, the researchers argue that AI-based auto-planning could help standardize care, particularly in clinics with limited resources or less specialized staff. They emphasize, however, that clinicians should review and adjust these automated plans before delivering treatment.

Future work will focus on testing the tool in real clinical settings, ensuring it works reliably across different hospitals, and adapting the approach to other cancer types, including breast and prostate cancer. For now, the work represents another step toward translating AI research in radiation oncology into practical tools, thanks to large-scale computing infrastructure like Voyager. Read the full paper at this link.

Related