# From data to treatment plan: An AI‐driven path for automated breast radiotherapy planning

## TL;DR

This paper introduces an AI system that automatically selects and generates high-quality radiotherapy plans for breast cancer, reducing planning time while maintaining clinical standards.

## Contribution

The novel contribution is TARS-B, an end-to-end automated framework for breast cancer radiotherapy planning that integrates AI-based decision-making and treatment plan generation.

## Key findings

- Automated plans showed no significant dosimetric differences compared to manual plans for both 3D-CRT and IMRT.
- TARS-B recommended re-planning 15 IMRT cases with 3D-CRT, with 14 meeting all clinical criteria.
- Planning times were significantly reduced, with IMRT plans dropping from 157 minutes to 2 minutes and 3D-CRT from 112 minutes to 5 minutes.

## Abstract

Breast cancer is one of the most prevalent malignancies in women, with radiotherapy (RT) playing a key role in its treatment. Advances in RT techniques, such as 3D conformal radiotherapy (3D‐CRT) and intensity‐modulated radiotherapy (IMRT), have improved dose precision and reduced side effects. However, RT modality selection and treatment planning remain manual, time‐consuming, and subject to variability.

This study presents and validates TARS‐B (Treatment Automation and Radiotherapy Selection for Breast Cancer), an automated framework that combines a deep learning‐based decision‐making module (DMF) for selecting the optimal RT technique and a fully automated treatment planning system (ATP) for generating deliverable plans that meet clinical quality standards and are deemed acceptable for clinical use.

TARS‐B functions in two stages. First, the DMF analyzes individual patient data to determine whether 3D‐CRT or IMRT is more appropriate. Second, the ATP generates the corresponding treatment plan. For 3D‐CRT, a field‐in‐field (FiF) method is used to enhance dose homogeneity and minimize hotspots. For IMRT, the DMF provides neural network‐based dose predictions, which are used to generate constraints for organs‐at‐risk (OARs). Both processes are fully scripted within the treatment planning system (TPS).

The framework was tested on 60 breast cancer patients: 30 originally treated with 3D‐CRT and 30 with IMRT. Two analyses were conducted. First, the ATP's performance was evaluated by comparing automated plans with their manually generated clinical counterparts for both techniques. Second, the full TARS‐B pipeline was assessed by applying the DMF to select the RT modality and automatically generating the plan, comparing results to the original clinical plans. Dosimetric parameters, including planning target volume (PTV) coverage, OAR constraints, and low‐ and intermediate‐dose bath, were analyzed. Planning times were also compared.

No statistically significant differences (p>0.005) were found between manual and automated plans in key dosimetric metrics, including PTV coverage (V95%), hotspots (V105%), and OAR constraints, for both 3D‐CRT and IMRT. TARS‐B confirmed the appropriateness of 3D‐CRT in all patients originally treated with it and recommended re‐planning with 3D‐CRT for 15 of 30 IMRT cases. Of these, 14 re‐plans met all criteria; one failed due to anatomical anomalies.

Re‐planning led to a reduction in low‐dose bath (up to 2800 cm3) and intermediate‐dose bath (up to 3000 cm3). The reduction in intermediate‐dose bath was statistically significant (p<0.005). Planning times decreased substantially: from 157.4±116.2 to 2.0±1.3 min for IMRT, and from 112±70 to 5±4 min for 3D‐CRT (p<0.005).

TARS‐B effectively automates both the selection of the most appropriate RT technique and the generation of high‐quality treatment plans. This framework improves workflow efficiency, reduces planning time, and preserves dosimetric quality, highlighting its potential for clinical implementation in breast cancer RT.