2912 - Early Prediction of Moderate-to-Severe Radiation-Induced Breast Damage using Ultrasound and Attention-Guided Deep Learning

Purpose/Objective(s):

Moderate-to-severe radiation-induced breast-tissue damage negatively affects cosmesis and quality of life after breast-conserving radiotherapy (RT), yet reliable early prediction remains limited. We hypothesize that early ultrasound-detected tissue changes reflect radiation response and can predict subsequent toxicity. We developed an attention-guided deep learning framework using longitudinal breast ultrasound (BUS) acquired pre-RT and end-of-RT to predict acute and late moderate-to-severe breast damage.

Materials/Methods:

A total of 151 breast cancer patients (median age 55) receiving breast RT underwent bilateral BUS at four quadrants (12, 3, 6, and 9 o’clock) pre-RT, end-of-RT, and at 6–8 weeks and 1 year post-RT. The primary endpoint was moderate-to-severe radiation-induced breast damage, defined on ultrasound as a BUS score =2 on a validated 0–3 scale reflecting changes such as edema and skin thickening at 6–8 weeks (acute) or 1 year (late) post-RT. Data were split at the patient level into training (60%), validation (20%), and independent test (20%) cohorts. For each quadrant, a ResNet-18 network extracted features from four input images (pre-RT and end-of-RT irradiated/normal pairs). A spatial attention mechanism dynamically weighted these inputs for quadrant-specific predictions. The validation-set Area under the Precision-Recall Curve (AUPRC) served as the primary model selection metric for class-imbalanced prediction, evaluating minority-class identification against prevalence-based baselines. Final performance is evaluated on the independent test set.

Results:

The independent test set included 31 patients (123 quadrants; acute prevalence 27.6%; late 22.7%). For acute toxicity, the model demonstrated robust predictive value (AUC 0.89, AUPRC 0.79 (baseline prevalence 0.28)). For late toxicity, it achieved an AUC of 0.75 and an AUPRC of 0.53 (baseline 0.23), well exceeding the prevalence-based baseline. Detailed performance metrics and 95% confidence intervals (CI) are in Table 1. Spatial attention analysis prioritized end-of-RT irradiated images, consistent with radiation-induced tissue changes.

Conclusion:

Attention-guided deep learning using longitudinal BUS acquired pre-RT and at end-of-RT demonstrates strong predictive performance for acute and late breast damage. The model enables early identification of patients at risk for moderate-to-severe toxicity despite low event prevalence. This objective imaging biomarker may provide a clinically actionable window for personalized surveillance and proactive toxicity management.

Table 1. Prediction performance (95% CI).

Metric	Acute Toxicity	Late Toxicity
AUC	0.89 (0.82-0.95)	0.75 (0.64-0.85)
AUPRC	0.79 (0.62-0.92)	0.53 (0.33-0.70)
Sensitivity	0.71 (0.56-0.86)	0.74 (0.57-0.90)
Specificity	0.92 (0.87-0.98)	0.62 (0.52-0.72)