After surgery for breast cancer, many patients receive radiation therapy to reduce the risk of the cancer returning. There are two main techniques used: 3D conformal radiation therapy (3D-CRT) and intensity modulated radiation therapy (IMRT). Each has advantages depending on the patient's anatomy.
Choosing between these techniques requires an experienced radiation oncologist or dosimetrist to evaluate each patient individually, often by creating a trial treatment plan for each option. This process is time-consuming and can vary between clinicians, leading to inconsistency.
This study developed a deep learning framework that can automatically evaluate both techniques and recommend the best one for each patient, streamlining the decision-making process without sacrificing accuracy.
The researchers trained a 2D U-Net convolutional neural network to predict the radiation dose distribution a patient would receive under each technique - without actually calculating the full treatment plan. The model learned from 346 breast cancer patients who had been previously treated.
A U-Net is a type of network architecture originally developed for medical image segmentation. It takes CT scan slices as input and outputs a map predicting how much radiation dose each tissue region would receive.
Two separate U-Net models were trained - one for IMRT and one for 3D-CRT - allowing the system to generate a predicted dose distribution for both techniques for any new patient and then compare them.
Once the AI predicts dose distributions for both IMRT and 3D-CRT, the framework evaluates them using clinical dosimetric criteria - rules defining how much dose is acceptable for the tumor target and surrounding healthy organs.
Key metrics evaluated include coverage of the clinical target volume (how well the tumor area receives the intended dose) and sparing of organs at risk such as the heart, lungs, and opposite breast.
The framework then applies a scoring system to each predicted plan and selects the technique that best balances tumor coverage with sparing of healthy tissue, mimicking how a clinician would weigh these trade-offs.
The decision-making framework achieved 90% accuracy in selecting the same radiotherapy technique that expert clinicians chose when they manually evaluated each patient. This demonstrates strong agreement between AI and clinical judgment.
Sensitivity for recommending IMRT when it was the correct choice was 95.7%, meaning the system rarely missed cases where IMRT was needed. This is clinically important because incorrectly using 3D-CRT when IMRT is needed could result in inadequate tumor coverage or excess organ dose.
The predicted dose distributions were also evaluated for accuracy. The AI-predicted doses closely matched the doses calculated by the actual treatment planning system, validating that the predictions were realistic enough to guide technique selection.
In clinical practice, generating even a rough comparison plan for each technique takes significant time. By automating technique selection, this framework could eliminate the need for manually testing both options for most patients.
The system is designed to integrate with existing clinical workflows. The CT scan data that is already collected for treatment planning can be fed directly into the AI, requiring no additional data collection or patient procedures.
Faster technique selection means patients can move through the radiotherapy preparation process more quickly, potentially reducing wait times before treatment begins.
This study demonstrates that a deep learning framework can reliably select between IMRT and 3D-CRT for breast cancer radiotherapy planning, achieving accuracy comparable to expert clinical decision-making.
The U-Net dose prediction approach is a key enabler - by predicting dose without running a full treatment plan, the system avoids the most time-consuming step while still capturing the information needed to make a good technique choice.
This framework represents a building block toward fully automated radiotherapy planning, where AI could handle routine decisions and allow clinicians to focus their expertise on complex cases.