Click here to

Session: Imaging Data Science for Treatment Assessment [Return to Session]

Ultrasound Radiomics-Based Machine Learning to Evaluate Radiation-Induced Acute Breast Toxicity in Breast Radiotherapy

J Wang*, b zhou, X Yang, J Lin, K Godette, S Kahn, M Torres, T Liu, Emory Univ, Atlanta, GA


TH-C-202-4 (Thursday, 7/14/2022) 10:00 AM - 11:00 AM [Eastern Time (GMT-4)]

Room 202

Purpose: To evaluate radiation-induced acute breast toxicity using ultrasound radiomics and machine learning (ML) models.

Methods: Seventy-two patients receiving breast radiotherapy (RT) were enrolled in a prospective, longitudinal study. All patients were scanned with breast ultrasound on 5 locations bilaterally, at 4 time points (pre RT, last day of RT, 3-5 weeks post RT, and 9-11 weeks post RT). In total, 2474 breast images were analyzed. Two regions of interest (ROIs) with 6 mm (shallow surface) and 12 mm depths were investigated. We extracted 163 radiomic features from the ROIs and trained 4 ML classifiers (KNN, avNNet, randomforest, and XGBoost) with five-fold cross validation on the training dataset (70% data), and then tested model performance on the remaining 30% data. Clinician graded toxicity scores were used as the ground truth: score 0 (none), 1 (mild), 2 (moderate), and 3 (severe toxicity).

Results: For binary classifications of with or without toxicity (score 0 vs 1, 2, and 3), we achieved the area under the receiver operating characteristics curve (AUC) of 0.88, accuracy of 0.81, sensitivity of 0.75 and specificity of 0.85. For binary classifications of mild or moderate/severe toxicity (score 0,1 vs 2,3), we achieved an AUC of 0.83, accuracy of 0.85, sensitivity of 0.92 and specificity of 0.62. For three-class classifications of none, mild and moderate/severe toxicity, we achieved an accuracy of 0.63, and weighted kappa of 0.53. For four-class classifications, we achieved an accuracy of 0.6, and weighted kappa of 0.49. In most cases, the choice of shallow surface ROI and avNNet worked best.

Conclusion: The current study explored the application of ML for auto-classification of radiation-induced acute breast toxicity. The radiomics-ML methods enable fast and quantitative evaluations of breast toxicity and could help to monitor the risk of developing severe toxicity during and after radiotherapy treatment.


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email