Toward Objective Grading of Reconstructed Breast Aesthetics: Three-dimensional Geometric and Color Features...

Interrater variance in subjective grading of reconstructed breast aesthetics limits the assessment reliability and reduces the quality of training datasets for machine learning models designed for objective, image-based grading. This study aims to identify the 3-dimensional (3D) features associated with such variability across 8 aesthetic viewpoints commonly used in Japan by analyzing case pairs in which expert scoring and pairwise comparisons show disagreement.

Two-dimensional (2D) and 3D images of the reconstructed breasts were obtained from 174 patients. Aesthetic outcomes were independently scored by expert raters using 2D images, and pairwise comparisons were conducted for 154 selected pairs. Several 3D geometric and color features were extracted from the 3D images; their correlation with expert scores was examined; and t tests were performed on the averaged feature values between the agreement and disagreement pairs under Benjamini-Hochberg false discovery rate correction.

Disagreement pairs were found irrespective of expert score differences. However, these pairs showed larger differences in the coefficient of variation between the winner and loser cases. For all the viewpoints, several 3D features revealed significant differences in their average values between agreement and disagreement pairs. These disagreement-specific features varied in both type and number across viewpoints and were different from the score-correlated features.

Viewpoint-specific 3D features associated with interrater variance were identified. This highlights a potential limitation of training machine learning models solely on expert scoring and suggests that integrating these 3D features may help inform future model design.