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

Sangjoon Park, Yong Bae Kim, Jee Suk Chang, Seo Hee Choi, Hyungjin Chung, Ik-Jae Lee, Hwa Kyung ByunarXiv 2024 · DOI · arXiv

Introduction

Post-surgical breast cosmesis—the aesthetic outcome following breast cancer surgery—is a critical determinant of a patient’s long-term quality of life. Clinical evaluation typically relies on the modified Harvard-Harris scale, ranging from "Excellent" to "Poor." However, this qualitative approach suffers from significant inter-observer variability. While semi-automated tools like BCCT.core offer more consistency, they require labor-intensive manual delineation of anatomical landmarks. To address these challenges, we present Attention-Guided Denoising Diffusion Anomaly Detection (AG-DDAD), a fully automated, unsupervised framework for objective cosmesis assessment.

The AG-DDAD Framework

The core innovation of AG-DDAD lies in its ability to learn the distribution of "normal" (high-quality) cosmetic outcomes and identify deviations as anomalies. The pipeline operates in two primary stages:

  1. Self-Supervised Masking: A Vision Transformer (ViT) trained via the DINO (Distillation with No Label) method generates attention maps. These maps naturally highlight semantically relevant regions—specifically the breasts—without requiring manual segmentation.
  2. Hybrid Diffusion Sampling: The model utilizes a diffusion process to reconstruct input images. In regions identified as potentially anomalous by the DINO mask, the system uses DDIM sampling to "re-imagine" the area as a healthy/normal version. Conversely, in "normal" regions, DDAD sampling preserves original fine details like skin texture and background.
Overview of the AG-DDAD pipeline, illustrating the forward diffusion, DINO-ViT attention masking, and the hybrid sampling process for reconstruction.

By training exclusively on images with "Excellent" or "Good" cosmesis, the model learns to reconstruct any input as an "idealized" version of itself. The reconstruction error—the difference between the original image and its normalized counterpart—serves as an objective anomaly score.

Experimental Setup

The model was trained on an unlabeled dataset of 1,237 photographic images from patients who underwent breast-conserving surgery. Evaluation was performed on a separate test set of 300 patients, categorized into four groups (Excellent, Good, Fair, Poor) based on the consensus of two board-certified radiation oncologists. Disagreements were resolved via consensus to create a robust ground truth benchmark.

Performance and Validation

AG-DDAD demonstrates state-of-the-art performance in identifying poor cosmetic outcomes, achieving an Area Under the Curve (AUC) of 0.921. This significantly outperforms other reconstruction-based anomaly detection methods such as standard DDIM (0.866) and PatchCore (0.710).

Furthermore, the model's anomaly scores show a clear, statistically significant correlation with clinical grades. As the cosmetic quality decreases from "Excellent" to "Poor," the anomaly scores systematically increase, providing a continuous metric for surgical success.

Box plot showing the distribution of anomaly scores across the four clinical cosmesis groups, demonstrating a significant upward trend as cosmetic quality declines.

The model also achieved a Cohen’s kappa of 0.60, matching the reliability of semi-automated tools like BCCT.core (0.67) but without the need for manual intervention.

Conclusion and Future Work

AG-DDAD provides a scalable, objective, and interpretable solution for breast cosmesis evaluation. By leveraging unsupervised learning and attention-guided diffusion, it eliminates the need for expensive manual labeling while providing visual "anomaly maps" that explain its scoring. Future research will focus on reducing processing time (currently ~15 seconds per image) and validating the model across diverse multi-institutional datasets to ensure robustness against varying photographic conditions.

Summary updated

BibTeX
@misc{park2024objective,
  author = {Sangjoon Park and Yong Bae Kim and Jee Suk Chang and Seo Hee Choi and Hyungjin Chung and Ik-Jae Lee and Hwa Kyung Byun},
  title = {{Objective and Interpretable Breast Cosmesis Evaluation with Attention Guided Denoising Diffusion Anomaly Detection Model}},
  year = {2024},
  eprint = {2402.18362},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2402.18362},
}
Figure