Related Experiment Video
Updated: Apr 28, 2026
Voluntary Breath-hold Technique for Reducing Heart Dose in Left Breast Radiotherapy
Published on: July 3, 2014
Machine Learning-based Prediction of Mean Heart Dose and Deep Inspiration Breath-hold Selection in Left-sided Breast
Deepali Patil1, Mukesh Kumar Zope1, Rishi Raj2
1Department of Medical Physics and State Cancer Institute, Indira Gandhi Institute of Medical Sciences, Patna, Bihar, India.
Background:
Deep inspiration breath-hold (DIBH) can reduce cardiac radiation exposure in left-sided breast cancer, but resource limitations necessitate appropriate patient selection.
Purpose:
To develop and evaluate a machine learning-based tool for predicting heart mean dose and identifying patients who would benefit from DIBH using simple anatomical predictors in left-sided breast cancer.
Materials And Methods:
A retrospective study analyzed 120 patients' treatment plans on free-breathing (FB) scans from left-sided postmastectomy breast cancer patients treated with volumetric modulated arc therapy (VMAT). All plans were generated using three techniques: VMAT 2-field plan (VMAT-2P), VMAT 4-field plan (VMAT-4P), and VMAT 5-field plan (VMAT-5P). Two anatomical predictors, maximum heart distance (MHD) and heart-to-PTV distance (HPD), were measured. Elastic Net regression was used for continuous dose prediction, whereas logistic regression was applied for binary classification of DIBH necessity, using a 5 Gy heart mean dose threshold. An independent cohort (n = 25) with paired FB-DIBH scans validated predictions.
Results:
In the validation cohort (n = 25), DIBH reduced mean heart dose by 34% (1.72-1.86 Gy, P < 0.001 for both techniques) and decreased high-risk patients (>5 Gy) by 69%-80%. Strong correlations were observed between FB predictions and DIBH-achieved doses for anatomical parameters and VMAT-2P (r = 0.667-0.720, P < 0.001), with moderate correlation for VMAT-4P (r = 0.545, P = 0.005). In the independent test cohort from the model development dataset (n = 24), Elastic Net achieved mean absolute errors of 0.81-1.02 Gy. Logistic regression demonstrated 87.5% accuracy with 83%-92% sensitivity and 83%-92% specificity for VMAT-2P and VMAT-4P (area under the curve [AUC]: 0.85-0.94). The VMAT-5P technique showed reduced classification performance (58.3% accuracy, AUC: 0.83).
Conclusions:
Machine learning software demonstrated accurate prediction of mean heart dose during pre-planning for left-sided breast cancer, enabling informed DIBH selection for cardiac sparing based on simple anatomical metrics from FB computed tomography (CT) scans.
