The research, presented at the international MEDICON 2026 congress in Siena and published by Springer Nature, analyses an Artificial Intelligence system for screening mammography. The study, authored by Luca Piano, Domenico Ficili, Giovanni Maimone, Giovanni Serra, Elena Galeasso, Giulia Rotunno, Simone Mazzetti, Valentina Giannini, Daniele Regge and Francesca Caumo, points to a possible 75% reduction in second-reader workload while maintaining high sensitivity.
Reducing radiologists’ workload without compromising the accuracy of screening mammography is one of the most important challenges for the application of Artificial Intelligence in healthcare. This is the focus of the study AI-Assisted Triage in Screening Mammography: Multi-Center Validation of a Transformer-Based System for Workload Reduction at High Sensitivity, presented by Giulia Rotunno, researcher at Health Triage, at MEDICON 2026, the international conference on medical and biological engineering held in Siena.
The paper was published on 26 September by Springer Nature in the IFMBE Proceedings series, volume 145, which collects the contributions to the congress. It presents the results of the evaluation of an AI system designed to analyse mammography exams and support the organisation of screening programmes.
More than 660,000 exams to develop the model
Screening mammography is a fundamental tool for the early detection of breast cancer, but it requires a considerable commitment of specialist resources. In many programmes, each exam is assessed independently by two radiologists, according to the principle of double reading, to increase the reliability of the diagnostic pathway.
The research addresses this need through an Artificial Intelligence-based triage system, developed to identify low-risk exams and help reduce the number of readings required. The model was trained on more than 660,000 mammography exams and then tested on two independent datasets, using images acquired with equipment from different manufacturers and exams performed on women of different nationalities. This approach made it possible to study the system’s ability to maintain high performance even under conditions different from those used during training — a particularly important aspect for potential applications of AI in heterogeneous clinical settings.
From a technological standpoint, the MedTinyViT AI model uses attention mechanisms that enable it to identify and relate complex information within images. Specifically, it analyses the four standard views of a mammography exam, integrating them into a single assessment.
The results: fewer exams requiring double reading
The study evaluated an Artificial Intelligence system designed to identify mammography exams that can be considered negative with a high degree of reliability, with the aim of reducing the number of exams requiring a second reading by a radiologist. In the evaluations carried out on two independent test sets, the model achieved a sensitivity of 98.9% and 96.6% respectively. This is a particularly important indicator for a triage system whose goal is to identify negative mammography exams, minimising the risk of excluding from second reading cases that instead require further investigation.
In the configuration where Artificial Intelligence is used to support radiologists, the study also found a reduction of approximately 75% in second-reader workload. The results, obtained under the conditions described by the researchers, show the potential of Artificial Intelligence to make screening mammography programmes more efficient, while maintaining high sensitivity and preserving the role of radiologists in clinical assessment.
Research combining technological and clinical expertise
The study is the result of a multidisciplinary collaboration between Health Triage researchers and professionals from leading Italian clinical and academic institutions: the Candiolo Cancer Institute – FPO IRCCS, the Department of Oncology of the University of Turin and the Breast Unit of Ospedale Pederzoli in Peschiera del Garda. The presentation at MEDICON 2026 and the subsequent publication in the congress proceedings represent an important opportunity to share the results with the international biomedical engineering community.
For Health Triage, this research is part of the development of Artificial Intelligence technologies applied to cancer prevention and diagnostics, with particular attention to scientific validation and the integration of these tools into clinical workflows.