Artificial intelligence in breast cancer imaging: a systematic review of diagnostic performance, predictive modeling, and clinical translation.

Across lesion detection and benign-versus-malignant classification tasks, deep learning models achieved AUCs of 0.80-0.95 across mammography, MRI, and ultrasound, whereas MRI-based treatment-response prediction models achieved AUCs up to 0.97.

Internally validated models consistently reported higher performance, indicating systematic performance inflation of approximately 5-15%.

Moderate to high risk of bias was observed, particularly in patient selection and applicability domains.