Systematic review of AI across breast cancer screening, pathology, radiomics and prognosis

Background

Breast cancer is the malignancy most often diagnosed in women worldwide, with approximately 2.30 million new cases and 666,103 deaths in 2022¹. This systematic review assessed the performance, rigour of methods and clinical applicability of AI primary studies in six areas: imaging for screening and diagnosis, digital pathology, radiomics, care around surgery, prognosis and natural language processing (NLP)¹.

Methods

The Cochrane Library, Embase, Web of Science, Scopus and PubMed/MEDLINE were searched from January 2015 to May 2026¹. Only original studies with patient-level or image-level data were included, and secondary syntheses were excluded¹. QUADAS-2 and PROBAST were used to judge risk of bias, and reporting followed PRISMA 2020¹.

Results

Of 5,275 records, 17 primary studies were eligible¹.

• Screening mammography: seven studies (7,540 to 144,231 examinations) reported deep learning AUC values of 0.876 to 0.939¹. In a randomised trial of 80,033 women, AI was non-inferior to double reading and cut radiologist workload by 44.3%¹.
• Digital pathology: five studies (270 to 44,732 whole-slide images) reported AUC 0.985 to 0.994, with deep learning outperforming pathologists working without assistance in detecting lymph node metastases¹.
• Radiomics: MRI-based radiomics achieved AUC 0.80 for predicting pathological complete response among patients with HER2-positive tumours¹.
• Prognosis: deep survival networks reached a concordance index of 0.626, compared with 0.612 for Cox regression¹.
• NLP: detection of metastatic recurrence in clinical records showed 83% sensitivity and 73% specificity¹.
• Perioperative care: none of the eligible studies addressed this area specifically in breast cancer¹.

Quality and limitations

Imaging and pathology studies carried low-to-moderate risk of bias, whereas small radiomics studies carried high risk¹. Only two imaging studies reported results stratified by breast density¹. Predicting receptor status from haematoxylin and eosin slides has not yet been shown to safely replace immunohistochemistry¹.

Authors' conclusions

The authors judged AI performance clinically meaningful in the five domains containing eligible studies¹. The lack of perioperative or intensive care studies specific to breast cancer was the largest evidence gap¹. Recurring limitations were external validation, diversity of demographics and evidence of prospective clinical integration¹. Receptor status prediction without immunohistochemistry needs thorough, prospective validation of its safety before it can be used alongside standard testing¹.