AI-Assisted Breast Cancer Screening: A Non-Inferiority Trial of Workload, CDR & RR

A new clinical trial from Spain is offering a glimpse into how artificial intelligence could reshape breast cancer screening, potentially reducing the workload for radiologists without compromising accuracy. The study, conducted at the Reina Sofía University Hospital in Córdoba, Spain, investigated whether an AI-assisted approach to reading mammograms and digital breast tomosynthesis (DBT) – often called 3D mammography – could streamline the process while maintaining high detection rates. The findings, currently available as a pre-print, suggest a promising path toward more efficient screening programs.

How the Trial Worked: A Paired Approach

The research team designed a paired, noninferiority trial, meaning each participant received two different reading strategies. The standard approach involved double reading by human radiologists – two radiologists independently reviewing each mammogram. The AI-assisted strategy likewise used double reading, but incorporated a commercially available AI system, Transpara (version 1.7, ScreenPoint Medical), to prioritize cases. Transpara flagged images with a higher probability of cancer (scoring 8 to 10 out of 10), and these were reviewed by two additional radiologists. Images with lower scores (1 to 7) were automatically classified as normal, effectively reducing the number of images requiring extensive human review. This approach aimed to reduce radiologist workload by focusing expertise on the cases most likely to require attention.

The trial enrolled 27,000 women aged 50 to 71, invited to participate in the Andalusian screening program in Spain between March 2022 and January 2024. All participants provided written informed consent, and the study received approval from the Reina Sofía University Hospital’s Institutional Review Board. The study design was preregistered at ClinicalTrials.gov, ensuring transparency and adherence to rigorous scientific standards.

AI’s Role: Identifying Suspicious Areas

The Transpara system uses deep convolutional neural networks – a type of machine learning – to analyze mammography images and identify areas that may be indicative of breast cancer. It provides radiologists with visual markings highlighting suspicious regions, along with a “suspiciousness score” indicating the likelihood of malignancy. The system is compatible with mammography equipment from major manufacturers like Siemens Healthineers, Hologic, and General Electric. Researchers have previously demonstrated that this AI system can achieve cancer detection performance comparable to radiologists alone, and can improve accuracy when used as a support tool. A 2019 study, for example, showed stand-alone AI performance rivaling that of 101 radiologists.

Key Outcomes: Workload, Detection, and Recall Rates

The primary outcomes of the trial focused on workload (the total number of radiologist readings), cancer detection rate (CDR – the number of cancers detected per 1,000 screenings), and recall rate (RR – the proportion of women called back for further evaluation after initial screening). Secondary outcomes included positive predictive value (PPV) of recalls (the proportion of recalled women who are ultimately diagnosed with cancer) and false positive rate (FPR – the proportion of recalled women who do not have cancer). The researchers are currently analyzing the data to determine whether the AI-assisted strategy is non-inferior to the standard approach in terms of CDR and RR, and whether it significantly reduces radiologist workload.

Understanding Noninferiority

It’s important to understand the concept of “noninferiority” in this context. The goal wasn’t necessarily to prove that AI *improves* cancer detection, but rather to demonstrate that it doesn’t *worsen* it while offering potential benefits like reduced workload. The researchers established a pre-defined margin of acceptable difference – a 5% relative reduction in CDR or a 5% relative increase in RR – to determine whether the AI-assisted strategy was noninferior. If noninferiority is established, they will then assess whether the AI strategy is superior in terms of workload reduction.

Safety and Data Integrity

The study protocol prioritized patient safety and data privacy. All mammographic images were anonymized before analysis, and data handling complied with applicable data protection regulations. The Institutional Review Board determined the study posed minimal risk to participants. No adverse events were reported during the trial. The researchers also emphasize that no clinical data used in the trial was used to develop the AI algorithm, mitigating potential bias.

What Comes Next: Data Analysis and Potential Implications

The research team is currently completing the statistical analysis of the data. The individual participant dataset and study protocol are publicly accessible via Zenodo at https://doi.org/10.5281/zenodo.17625633, promoting transparency and allowing for independent verification of the findings. If the results confirm that the AI-assisted strategy is noninferior and reduces workload, it could have significant implications for breast cancer screening programs worldwide. It could help address radiologist shortages, reduce burnout, and potentially improve the efficiency of screening services. However, it’s crucial to remember that this is just one study, and further research will be needed to validate these findings in different populations and settings. The Office for Human Research Protections (OHRP) provides oversight for ethical research practices, ensuring the safety and wellbeing of participants in studies like this one.

the integration of AI into breast cancer screening is likely to be a gradual process, requiring careful evaluation and ongoing monitoring to ensure that it benefits both patients and healthcare providers. The results of this Spanish trial represent an important step forward in understanding the potential of AI to improve the effectiveness and efficiency of this vital screening program.