A comprehensive study recently published in JAMA Network Open has investigated the efficacy of an artificial intelligence (AI) algorithm in predicting breast cancer through mammogram screenings. The study, which was conducted as a population-based, retrospective cohort analysis, utilised data from BreastScreen Norway. This national screening programme caters to 680,000 Norwegian women aged between 50 and 69 years, offering biennial digital mammography screening.
The research aimed to ascertain whether a commercially available AI algorithm, named INSIGHT MMG, could accurately estimate the risk of future breast cancers being diagnosed in subsequent screening rounds. A total of 246,472 women participated, contributing to 671,828 mammogram screenings between 2017 and 2021. Importantly, participants in the study did not have a prior history of breast cancer and had undergone at least three consecutive screenings, where the first two did not involve any cancer diagnosis.
In alignment with conventional medical assessments, the study utilised a score-based system to classify findings on a scale from 1 to 5. A score of 1 indicated normal findings, whereas a score of 5 suggested a high suspicion of malignant neoplasm. An AI-generated cancer detection score complemented these classifications, ranging from 0 to 100, with higher scores signalling a greater likelihood of cancer presence.
During the study period, approximately 1265 cancers were detected via screening, and 342 interval cancers were diagnosed. The mean age of the participants at diagnosis was around 58.5 years for screen-detected cancers and 57.4 years for interval cancers, while the mean age for women without breast cancer was 56.4 years.
The algorithm displayed increased accuracy in predicting breast cancer as the screening rounds progressed. For instance, mean AI scores for breasts that developed screening-detected cancer increased markedly from 19.2 at the first screening round to 82.7 at the third. In contrast, those that did not develop cancer exhibited a decline in scores from 9.5 to 5.0 over the same period. The algorithm's efficiency was notably demonstrated through the area under the receiver operating characteristic curves (AUCs), which improved from 0.64 to 0.97 across the three screening rounds for detecting screen-diagnosed cancer.
However, the report acknowledges several constraints within the study's design. As it was based on retrospective data, the study evaluated only one AI detection system and the participants were predominantly White women. This raises questions about the applicability of the findings to broader, more diverse populations. Future research should aim to validate these findings across a range of AI tools and patient demographics to ensure inclusivity in breast cancer forecasting technology.
The retrospective nature of the study provides a robust dataset, yet highlights the need for broader multi-system evaluations. As AI continues to integrate into medical diagnostics, this research showcases the potential for AI algorithms to enhance predictive accuracy in breast cancer detection, potentially paving the way for the adoption of smarter, data-driven approaches to mammography screenings.
Source: Noah Wire Services