Serial skin photographs predict significant breast radiation dermatitis by treatment week three

A dual-view AI model predicted grade ≥2 breast radiation dermatitis with AUC 0.91 overall and approximately 0.80 by week three.

KEY POINTS

  • This prospective single-center observational study enrolled 626 patients with breast cancer receiving postoperative whole-breast or chest wall radiotherapy between April 2024 and December 2025. Median age was 49 years; 277 patients (44.3%) developed RTOG grade ≥2 radiation dermatitis, including 253 grade 2 and 24 grade 3 events.
  • Standardized photographs were acquired from two anatomical views, covering the breast/chest wall and axillary region, at predefined intervals during treatment. The model combined EfficientNetV2-L for image feature extraction with a gated recurrent unit to capture longitudinal skin changes, using patient-level five-fold cross-validation for internal evaluation.
  • The combined spatial–temporal model achieved an AUC of 0.9129 (95% CI 0.8893–0.9345) for predicting grade ≥2 dermatitis. At the selected threshold, sensitivity was 85.6%, specificity 83.9%, overall accuracy 84.6%, and 237 of 277 patients who developed clinically significant dermatitis were correctly identified.
  • Both anatomical views and temporal information contributed to discrimination. Removing axillary images reduced AUC from 0.9129 to 0.8704, while replacing longitudinal analysis with a static first-week model reduced AUC to 0.6997, indicating that evolving skin appearance carried substantial predictive information.
  • Prediction accuracy improved progressively as additional images became available. Using only the first three treatment weeks, the model achieved an AUC of approximately 0.80, rising to 0.8931 at week five. The model therefore showed potential for identifying patients at risk before later, more pronounced skin reactions developed.
  • Clinical event rates varied across treatment schedules: grade ≥2 dermatitis occurred in 55.6% of patients receiving 50 Gy in 25 fractions, compared with 17.7% receiving 40.05 Gy in 15 fractions followed by 10.68 Gy in four fractions. These observational differences were not adjusted comparisons of fractionation strategies and should not be interpreted causally.
  • Interpretation is limited by single-center internal validation, controlled photography conditions, manually selected skin regions and limited representation of diverse skin pigmentation. Only 291 patients had complete imaging through week five, and 39 through week six; no independent external cohort or prospective intervention study demonstrated that AI-guided skin management reduces dermatitis.

CLINICAL TAKEAWAY

Longitudinal photographs of the breast and axilla may provide a relatively accessible approach to identifying patients at risk of clinically significant dermatitis during radiotherapy, potentially supporting earlier monitoring and supportive care. However, the strong cross-validation results do not establish real-world performance across different cameras, lighting conditions or patient populations, and the impact of risk-based interventions remains untested.

SOURCE

Radiation Oncology

Browse more research Suggest a correction