Table 1.

The application of AI in early detection and diagnosis of BC.

Study & cohort	Model/method	Key results	Ref.
U-Net + CBR; MG images	Segmentation + case-based reasoning	Malignancy classification Acc 91.34%; segmentation Acc 86.71%.	(28, 29)
Hybrid CNN features + variance thresholding (TV) + MSVM; mini-MIAS	ResNet50/AlexNet features + MSVM	Acc 99% on mini-MIAS (small curated dataset).	(31)
Cross-channel convolutional correlation features	Deep learning	Recall 95.65%.	(32)
Peritumoral GLCM/gradient + contralateral normal-breast features	Hand-crafted textures	AUC improved 0.79 → 0.84.	(33)
CEM radiomics with auto-segmentation and tri-compartment quantification	Radiomics model	PPV 0.49 (+47% vs MG); non-essential biopsies −35.8%.	(34, 35)
Multi-sequence MRI radiomics + BI-RADS	Radiomics + BI-RADS	AUC 0.95 for HER2; AUC 0.98 for diagnosis when combined with BI-RADS.	(36)
AI-STREAM (Korea; n=24,543)	AI-CAD assisting readers	For subspecialists, CDR + 13.8% (5.70‰ vs 5.01‰, p<0.001) with no RR increase (4.53% vs 4.48%); stand-alone AI CDR 5.21‰, RR 6.25%.	(37)
Multi-center 5,025 pts; prospective 187 pts)	Multimodal ML (BMU-Net framework)	Prospective Accuracy 90.1%; comparable to experts for benign–malignant, superior for fine-grained pathology; close to pre-biopsy pathology 92.7%.	(38)