*RF01-09 BI-RADS mammograms classification methodology Details

Start Date: 2025-09-20

End Date: 2025-09-20

We present a methodological framework for automated BI-RADS categorization in mammography that prioritizes rigor, scalability, and clinical applicability over hyperparameter optimization. The approach consists of: (i) object detection to isolate the breast region using a modern YOLO family detector, enabling automatic dataset cleaning and standardized cropping; (ii) cloud-based workflows to ensure reproducibility, collaboration, and auditable experiment tracking; (iii) principled data reorganization and inspection to address class imbalance across ordinal BI-RADS categories; (iv) strong data augmentation to emulate clinical variability; and (v) two-phase training with transfer learning—first adapting the classification head, then gradual fine-tuning—using efficient convolutional architectures inspired by the MobileNet philosophy (accuracy–efficiency trade-off suitable for constrained settings). Evaluation emphasizes per-class metrics, confusion matrices, and ordinal-aware agreement measures, coupled with model versioning and single-case testing to reflect practical usage. Rather than reporting peak scores, we articulate a generalizable method aligned with emerging reporting guidelines for AI in medical imaging. This framework is intended as a reproducible baseline for prospective, multi-center validation and for integration into screening workflows, especially in resource-limited environments.