A clinical trial involving more than 100,000 women shows that AI improves breast cancer screening

The MASAI trial shows notable strengths in that it is a randomised, prospective, large-scale clinical trial (>100,000 women) implemented in a real-world setting and compared with double reading by radiologists. All of this gives it a higher level of evidence than previous technical validation studies.

From a breast radiology perspective, it analyses whether the incorporation of AI into screening improves breast cancer detection in a way that reduces interval cancers and detects less aggressive tumours, going beyond traditional indicators such as detection and number of recalls for complementary studies.

It stands out for evaluating clinically relevant objectives in a population-based programme, such as the rate of interval cancers, tumour stage, biological aggressiveness and impact on workload.

Radiological results show that AI provides a clinically relevant improvement to mammographic screening with a 12% reduction in interval cancers; this is one of the main weaknesses of conventional screening due to the greater aggressiveness and worse prognosis of this type of tumour. The reduction in the number of invasive, larger and aggressive subtypes of tumours indicates that AI not only detects more cancers, but detects them earlier and from a more favourable biological point of view, probably thanks to better identification of subtle lesions that are difficult to assess with the human eye. All this is achieved without an increase in false positives, maintaining rates of around 1.5%, which avoids overloading the system, reduces patient anxiety and reinforces the idea that AI acts as a specific discrimination tool and not merely as an amplifier of suspicions.

The trial stands out for its integration of AI into the workflow, using it for the triage of low-risk cases, focused double reading of high-risk cases, and as a decision support tool, not as a substitute for the radiologist. From a clinical perspective, the 44% reduction in the reading workload is highly relevant, as it frees up time for complex cases and addresses a key structural problem, namely the shortage of radiologists with expertise in breast imaging. In this sense, the study does not propose replacing the radiologist, but rather redefining the organisation of work in a realistic and acceptable way.

The study has several limitations, as it was conducted only in Sweden, with a specific healthcare system and population, evaluating a single AI system and type of mammography machine, and with radiologists of moderate to high experience. All of this could restrict the generalisation of the results to other contexts, technologies, or levels of experience. Data on race or ethnicity were also not included, preventing analysis of AI performance in diverse populations, and although follow-up was two years, long-term outcomes and the cost-effectiveness of AI use were not evaluated.

Overall, if its effectiveness is confirmed, this study could drive structural change in breast cancer screening programmes, accelerating the responsible integration of AI into routine clinical practice.

Furthermore, if the effectiveness of AI in mammography screening is confirmed, its implications would be far-reaching. Clinically, it would allow for earlier detection of tumours with a reduction in interval cancers and a potential decrease in mortality.

For healthcare systems, it would improve efficiency and alleviate the workload of radiologists. Economically, it could reduce the costs of advanced treatments, although it would require initial investment. Organisationally and professionally, it would transform the role of the radiologist and require specific training in AI. Ethically, it would demand clear regulatory frameworks, bias monitoring and transparency to build trust.

Overall, it could drive structural change in screening programmes and the responsible integration of AI into clinical practice, not as a replacement for radiologists but as initial triage and reinforcement in complex cases.