3423 - Objective and Interpretable Breast Cosmesis Evaluation with Attention Guided Denoising Diffusion Anomaly Detection Model

S. Park¹, Y. B. Kim², J. S. Chang³, S. H. Choi⁴, H. Chung⁵, I. J. Lee⁶, and H. K. Byun⁷; ¹Department of Radiation Oncology, Yonsei Cancer Center, Yonsei University College of Medicine, Seoul, Korea, Republic of (South), ²Yonsei Cancer Center, Yonsei University College of Medicine, Seoul, Korea, Republic of (South), ³Department of Radiation Oncology, BC Cancer, Vancouver, BC, Canada, ⁴Department of Radiation Oncology, Yongin Severance Hospital, Yonsei University College of Medicine, Gyeonggi-do, Korea, Republic of (South), ⁵Korea Advanced Institute of Science and Technology, Daejeon, Daejeon, Korea, Republic of (South), ⁶Yonsei Cancer Center, Seoul, Korea, Republic of (South), ⁷Department of Radiation Oncology Yongin Severance Hospital, Yonsei University College of Medicine, Yongin, Korea, Republic of (South)

Purpose/Objective(s): The assessment of cosmetic outcomes post-breast cancer surgery is critical for patient quality of life, yet it presents challenges due to subjective expert evaluations. This study introduces an automated, objective approach leveraging the Attention-Guided Denoising Diffusion Anomaly Detection (AG-DDAD) model to evaluate breast cosmesis, aiming to overcome limitations of conventional supervised learning and existing anomaly detection methods. Materials/

Methods: AG-DDAD employs a self-supervised Vision Transformer (ViT) for attention mask generation, combined with a diffusion model for high-quality image reconstruction. Trained on unlabeled data primarily representing normal cosmesis, this unsupervised method automatically scores breast cosmesis. The models effectiveness was validated against a curated dataset of 300 patients, comparing its performance with traditional rule-based programs and state-of-the-art (SOTA) anomaly detection models.
Results: The AG-DDAD model demonstrated a distinct correlation between anomaly scores and breast cosmesis quality, showing a clear progression in median scores from 6.628 (IQR: 6.119-7.412) for excellent to 9.052 (IQR: 8.001-10.520) for poor cosmesis. This confirms its precision in categorizing cosmesis levels. Outshining state-of-the-art (SOTA) anomaly detection models, AG-DDAD emerged as superior, highlighting its cutting-edge accuracy. It uniquely modifies images to enhance cosmesis while preserving non-relevant details, distinguishing it from models that either misapply changes or overlook key adjustments. Despite a slightly lower Cohen’s kappa value of 0.60 (0.53-0.67) compared to the BCCT.core programs 0.67 (0.50-0.72), AG-DDAD validates its capacity for objective, manual-free cosmesis evaluation. Offering a continuous, quantifiable assessment, it advances beyond conventional four-class systems.
Conclusion: AG-DDAD represents a significant advancement in the objective and interpretable evaluation of breast cosmesis post-surgery. By addressing the challenges of subjective expert assessments and the limitations of current anomaly detection models, this research paves the way for broader applications in unsupervised anomaly detection for various applications in radiation oncology. Future work will focus on external validation and the exploration of acceleration techniques to enhance the models clinical applicability. Abstract 3423 – Table 1: Performance comparison with SOTA anomaly detection models

Methods	AUC (95% CI)
Feature-based
PaDiM	0.556 (0.448-0.664)
CSFlow	0.644 (0.546-.0743)
PatchCore	0.710 (0.619-0.800)
Reconstruction-based
Ganomlay	0.730 (0.642-0.817)
DRAEM	0.676 (0.581-0.770)
DDPM	0.815 (0.742-0.887)
Proposed (AG-DDAD)	0.921 (0.876-0.965)