Exploratory Evaluation of Radiology Report Embeddings and Mammographic Radiomics for Post-Cryoablation Biopsy and Biopsy-Proven Ipsilateral Breast Tumor Recurrence.
Authors
Affiliations (2)
Affiliations (2)
- Department of Radiology, Brown University Health and The Warren Alpert Medical School, Brown University, Providence, RI, 02903, USA.
- Department of Radiology, Brown University Health and The Warren Alpert Medical School, Brown University, Providence, RI, 02903, USA. Electronic address: [email protected].
Abstract
Post-cryoablation surveillance is challenging because benign treatment-related imaging changes can overlap with findings concerning for residual or recurrent malignancy. We explored whether mammographic radiomics, radiology report embeddings, or multimodal fusion were associated with subsequent biopsy and biopsy-proven IBTR. This retrospective proof-of-concept study included 40 unique patients treated with breast cryoablation. One index pre-cryoablation mammogram and corresponding report were used per patient. Radiomics, BioClinical-ModernBERT-base report embeddings, and multimodal fusion were evaluated with a fixed L2-regularized logistic regression classifier using patient-level leave-one-out cross-validation. All learned preprocessing and feature-selection steps were fit within training folds. Repeated stratified 5-fold cross-validation assessed stability. Thirteen patients underwent post-cryoablation biopsy and 7 had biopsy-proven IBTR. For IBTR, ROC AUC was 0.42 (95% CI, 0.17-0.69) for radiomics, 0.53 (0.25-0.79) for report text, and 0.43 (0.15-0.73) for fusion. For biopsy, ROC AUC was 0.72 (0.53-0.87), 0.49 (0.30-0.69), and 0.85 (0.72-0.97), respectively. Fusion biopsy performance decreased in repeated 5-fold cross-validation (mean ROC AUC, 0.69 ± 0.06). No modality demonstrated detectable discrimination for biopsy-proven IBTR. Fusion yielded the highest observed leave-one-out discrimination for biopsy, but lower repeated-cross-validation performance precludes claims of superiority. Findings are hypothesis-generating and require larger cohorts, fixed follow-up horizons, and independent validation.