Bm1.0: A Radiomics Computational Dictionary Bridging Ai Research and Clinical Practice in Breast Cancer Care Publisher



Gorji A ; Sanati N ; Pouria AH ; Mehrnia SS ; Hacihaliloglu I ; Rahmim A ; Salmanpour MR
Source: Progress in Biomedical Optics and Imaging - Proceedings of SPIE Published:2026

Abstract

Artificial intelligence (AI) models for breast cancer diagnostics show great promise, but their “black box” nature limits clinical trust and adoption. This is especially true for radiomics features (RFs), mathematically derived from images yet disconnected from BI-RADS, creating an interpretability gap where AI outputs are opaque and non-actionable. In collaboration with medical professionals and AI engineers, we developed a dual-dictionary framework that translates abstract RFs into clinically meaningful terms, linking AI outputs to BI-RADS descriptors. By mapping hidden biomarkers to recognizable traits-tumor shape, margins, and internal enhancement, validated by three physicians, three physicists, and a radiologist, the dictionary provides a trusted, evidence-based interpretive baseline, enhancing explainability and clinical trust. We introduced a dual-dictionary framework consisting of an Expert-Informed Feature Interpretation Dictionary (ExpertFID), which maps 56 RFs to BI-RADS descriptors using literature and physician consensus, and an Automated Feature Interpretation Dictionary (AutoFID), which uses SHAP values from machine learning models to assign clinical meaning to 52 previously unmapped features. We validated this framework by training and testing 20 model pipelines on a multi-institutional, large cohort of 1,549 breast MRI cases to classify triple-negative (TNBC) versus non-TNBC subtypes. Our top-performing model (Extra Trees Importance feature selector + HistGradientBoosting classifier) achieved a robust validation accuracy of 0.75±0.05 on this large, diverse dataset, demonstrating strong generalizability. The Breast-MRI dual-dictionary framework (BM1.0) successfully mapped the most predictive quantitative imaging features into clinically interpretable descriptors, such as 'GLCM Correlation' and 'shape Flatness', into clinically intelligible concepts related to tumor shape and internal enhancement, aligning with known imaging phenotypes. The framework validates existing literature and offers data-driven insights into novel biomarkers. It provides a scalable, transparent methodology for interpreting radiomics models by linking features to the BI-RADS lexicon, bridging the gap between AI outputs and clinical practice, enhancing trust, and supporting AI integration in precision oncology. © 2026 SPIE. All rights reserved.