Multi-class Identification and Quantification of Breast Implant-related Complications in Ultrasonography Using a Pathology-informed Lesion Graph: A Deep Learning Development and Validation Study.

BACKGROUND: Breast ultrasonography (US) is cost-effective for detecting breast implant-related complications, but its reliability remains highly operator-dependent. OBJECTIVES: This study developed and validated a deep learning model for the automated multi-class identification and quantitative evaluation of breast implant complications (rupture, capsular contracture, and hematoma). METHODS: We retrospectively analyzed 14,679 breast US images across 1,022 breast implant sequences from 908 patients. The proposed architecture integrates multiscale visual encoding, automated probe position mapping, and a pathology-informed mask graph layer with a Gradient Reversal Layer for lumen adversarial debiasing. Four convolutional backbones were benchmarked using an independent patient-level test dataset (N=197 implants). RESULTS: The optimized DenseNet-based framework achieved 96.97% balanced accuracy with expert-guided spatial priors and 88.67% under autonomous baseline conditions. The model yielded 100% sensitivity for normal, rupture, and hematoma cohorts, and 87.88% sensitivity for capsular contracture; minor false-negatives were traced to morphological heterogeneity and coexisting pathologies. Furthermore, the system extracted quantitative morphometric profiles, yielding a mean relative capsule thickness of 2.4±0.4% for capsular contractures and a mean relative fluid area of 13.9±6.4% for hematomas. CONCLUSIONS: The proposed multiscale model delivers a comprehensive, clinical-grade output, comprising accurate complication classification, five-zone spatial localization, and objective morphometric metrics, while mitigating operator dependency. This system holds strong potential as an accessible, high-performance decision support tool for long-term safety screening of patients with breast implants.