# AI-enabled POCUS for breast cancer risk stratification in a resource-limited tertiary clinic

## TL;DR

This study evaluates an AI-powered ultrasound system for breast cancer risk assessment in a South African hospital, showing moderate accuracy in predicting malignancy.

## Contribution

The study introduces a locally developed AI-enabled POCUS system for breast cancer risk stratification in resource-limited settings.

## Key findings

- Breast AI achieved a sensitivity of 67.2%, specificity of 79.4%, and PPV of 70.3% at a 51% threshold.
- A three-tiered risk model (<30%, 30–51%, >51%) was developed based on AI scores for triaging suspicious breast lesions.

## Abstract

Breast cancer remains a major public health burden in South Africa, with diagnostic delays contributing to poor outcomes. Ultrasound is effective for early detection but is limited by access and operator variability. Integrating artificial intelligence (AI) into point-of-care ultrasound (POCUS) offers a potential solution.

To evaluate the diagnostic performance of a locally developed AI-enabled POCUS system (Breast AI) in predicting malignancy among women with palpable breast abnormalities.

A prospective cohort study was conducted between June 2024 and November 2024 at Groote Schuur Hospital. Women aged ≥ 25 years with suspicious breast lesions underwent Breast AI ultrasound prior to biopsy. Real-time malignancy risk scores were compared with histopathological results. Diagnostic accuracy was assessed using sensitivity, specificity, positive predictive value (PPV), F1 score and area under the curve (AUC).

Among 159 participants, Breast AI achieved a sensitivity of 67.2%, specificity of 79.4% and PPV of 70.3% at a 51% threshold. The AUC was 0.76, reflecting moderate discriminatory performance. F1 score analysis identified 51% as the optimal cut-off (F1 = 65.7%). Benign pathologies such as fibroadenomas and fat necrosis correlated with low AI scores. A three-tiered risk model was developed: < 30% (low), 30% – 51% (intermediate) and > 51% (high risk).

Breast AI demonstrates promising diagnostic accuracy for triaging suspicious breast lesions, particularly in resource-constrained settings.

This study provides real-world evidence supporting the integration of AI into POCUS to improve breast cancer detection and clinical decision-making in low-resource environments.