Artificial intelligence in healthcare: how diagnosis, imaging and clinical practice are changing

This framework is particularly important in diagnostics. Contemporary medicine produces volumes of data that no healthcare professional can manage without advanced digital tools. A multilayer CT scan can produce hundreds or even thousands of images; MRI examinations generate complex image sequences; and digital mammography, especially when associated with tomosynthesis, requires the evaluation of increasing volumes of information. In addition to these data, there are electronic health records, laboratory results, genomic data, digital pathology findings, and remote patient monitoring data. In this scenario, the challenge is no longer simply to collect clinical information but to interpret it accurately, efficiently, and sustainably.

In Italy, this challenge may be even more relevant, as several speakers highlighted during the forum. As Bechini explains, "Our patients are becoming increasingly complex due to demographic trends. We must manage this growing complexity by adopting and effectively governing AI technologies".

AI is becoming increasingly important because it can identify complex patterns within large datasets and apply analytical criteria consistently across different cases. This is particularly useful in high-volume contexts, where organizational pressure increases the risk of delays, interpretive variability, and cognitive overload.

The issue is far from theoretical. A study published in BMJ Quality & Safety in 2014 estimated that diagnostic errors affect about 5% of adults treated annually in outpatient settings in the United States, equal to at least one in twenty patients. Subsequently, the National Academies report, Improving Diagnosis in Health Care identified diagnostic error as a significant yet often underestimated patient safety concern.

Artificial intelligence does not eliminate the risk of diagnostic error, but it can help address some of the factors that contribute to it. Algorithms can systematically analyze medical images, flag suspicious findings, quantify lesions, compare current and previous examinations, prioritize urgent cases, and support physicians in generating more structured reports.

This role becomes particularly relevant given that the literature suggests the comparison should not be framed as "human versus machine," but rather as traditional workflow versus AI-assisted workflow. A systematic review and meta-analysis published in 2019 in The Lancet Digital Health showed that, in specific medical imaging tasks, certain deep-learning systems can achieve performance comparable to that of specialists. However, the authors also emphasized the need for rigorous validation and studies conducted under real-world clinical conditions. 

Radiology is the field in which this transformation is most advanced. Not surprisingly, radiological imaging generates highly information-rich digital data, and many radiological activities involve recognizing visual abnormalities, measuring lesions, and comparing subsequent examinations. Algorithms are now being used or tested to detect pulmonary nodules, intracranial hemorrhages, pulmonary embolisms, pneumothorax, fractures, breast lesions, and other abnormalities that require prioritization. In this context, AI can act both as a support for diagnosis and as a worklist organization tool, bringing potentially critical examinations to the radiologist's attention first.

Cardiology also offers strong examples of integrating AI into clinical workflows. A randomized, blinded trial published in 2023 in Nature (Blinded, randomized trial of sonographer versus AI cardiac function assessment) compared an initial assessment of left ventricular ejection fraction performed by AI with that performed by a sonographer in echocardiography. The study demonstrated the non-inferiority of the AI-assisted approach within the image-interpretation workflow, with a final review always entrusted to the cardiologist. It is a very clear example of the more realistic role of AI: not autonomous diagnosis, but quantitative pre-analysis that the physician verifies, corrects, and integrates into the overall clinical assessment.

Neurology, especially in time-dependent diseases, highlights another type of benefit: not only diagnostic accuracy, but also faster progression through the diagnostic pathway. Here the benefit to the patient is intuitive: if a critical case is recognised earlier, the team activates in a timely manner; if the patient is referred to the appropriate centre earlier, they can access treatment more quickly. Of course, even in this area, AI does not replace the clinical network. It only works if integrated with emergency departments, neuroradiology, neurology, hemodynamic/intervention, communication systems and territorial protocols.

AI is also being integrated directly into the equipment and software used for the acquisition, reconstruction, and quality control of magnetic resonance images. In this case, the value is not only in the automated interpretation of the completed examination, but in the possibility of obtaining more stable images, reducing acquisition times, limiting artifacts and making examinations more reproducible despite variations in protocols, equipment, and operator expertise.

In MRI, deep-learning reconstruction systems generate images from subsampled or noisy data, with the goal of speeding up the examination without compromising diagnostic quality. A study published in 2024 in Radiology: Artificial Intelligence reported that, following the implementation of deep-learning reconstruction methods in a multicenter clinical network, scan times were reduced by up to 53% and examination-room occupancy times by up to 41% across several examinations, including brain, lumbar spine, knee and shoulder. It is also a relevant point for orthopedics and sports medicine, where MRI and musculoskeletal imaging play a central role in evaluating joint, tendon, muscle and post-traumatic injuries.

In ultrasound, AI primarily supports the acquisition and standardization of a highly operator-dependent examination. In musculoskeletal ultrasound, instruments capable of guiding image acquisition and assessing image quality can reduce the variability between operators and improve the reliability of the acquired data. However, it remains important to distinguish technical benefits, such as reduced examination times and improved image quality, from clinical outcomes, including diagnostic performance, patient outcomes, and care-pathway organization.

In urology, AI is being integrated into stages of the diagnostic process where outcomes depend on image quality, targeting precision, and procedural reproducibility. An example comes from a prospective study by Oderda et al., published in Cancers in 2025, involving 148 patients undergoing transperineal prostate biopsy with MRI-ultrasound fusion using UroFusion, a system equipped with AI-guided self-contouring. The software automates prostate segmentation and alignment between ultrasound and magnetic resonance images, reducing operator-dependent variability.

In the study, the median time to obtain the fusion image was 5 minutes, the biopsy lasted 15 minutes and the tumour detection rate was 64%, with clinically significant tumours in 56% of cases. An important finding was that the detection did not differ between experienced operators and less experienced supervised operators. Systematic sampling detected additional tumours beyond the targeted areas, confirming that AI improves workflow, but does not eliminate the need for a comprehensive diagnostic strategy.

One of the most underestimated aspects of artificial intelligence in healthcare concerns the organization. AI can help organizations make better use of existing resources, including healthcare professionals, equipment, and clinical capacity. A system that identifies high-priority examinations can reduce the time before reading; an algorithm that automates repetitive measurements can free up clinicians’ time; software that structures information can improve communication between specialists. 

For the patient, this can translate into faster diagnosis, earlier access to the specialist, and a more timely start of therapy. The aim is not to claim that AI alone can eliminate waiting lists, because waiting times also depend on workforce availability, infrastructure, health planning and demand for services. The point is more precise: in well-designed workflows, AI can help alleviate specific bottlenecks within the care pathway, especially in the triage, pre-read, measurement, and prioritization phases.

Diagnostic imaging technologies are benefiting from AI throughout the process, not just in the final readout. Before image acquisition, algorithms may support protocol choice or positioning. During image acquisition, they can contribute to image reconstruction, noise reduction, quality optimization, and, in some applications, reduced scan or dose times. After acquisition, they can segment organs and lesions, calculate volumes, compare previous exams, and generate quantitative metrics useful for follow-up.

Radiomics, i.e. the extraction of quantitative features from medical images, makes it possible to transform the medical image into a source of measurable data, potentially integrated with biomarkers, clinical data and genomic information. However, the clinical value of this approach depends on validation: it is not enough to demonstrate that a model works on a retrospective dataset, it is necessary to demonstrate that it improves clinical decision-making, care pathways, and outcomes in real-world populations.

Precisely for this reason, the integration of AI into clinical workflows requires strict conditions. The first requirement is interoperability: an algorithm that is not interoperable with PACS, RIS, electronic medical records and hospital systems risks adding complexity instead of reducing it. The second requirement is clinical validation on representative populations, because a model trained on data from a certain context may perform worse when applied to different patients, equipment or protocols. The third requirement is monitoring over time: algorithm performance may decline over time if the characteristics of the input data change, a phenomenon known as dataset shift.

“It's important to develop competencies in this area”, says Bechini. “The real challenge isn't just technological: it's about professionals' ability to use it critically and consciously”, he adds.

"Rather than focusing on new professional roles (which are certainly needed), I would emphasize the transformation of existing ones. For this reason, it is necessary to integrate training on the use, limitations, and risks of AI into medical and healthcare education programs, rather than responding only after unintended consequences emerge", continues Bechini.

On this point, the European regulatory framework is particularly relevant. The European regulation on artificial intelligence, AI Act (Regulation (EU) 2024/1689), establishes specific rules for high-risk systems, with the aim of protecting health, safety and fundamental rights. The European Commission and the Medical Device Coordination Group have also published documents on the interaction between AI Act, Medical Devices Regulation and In Vitro Diagnostic Medical Devices Regulation, highlighting the complexity of classifying healthcare software and AI systems used in clinical practice.

The theme of trust is therefore central. Physicians and patients must know when an AI system has been used, for what purpose, with what limits and who retains professional responsibility for its use. Transparency does not necessarily mean making every mathematical parameter of a neural network understandable, but ensuring traceability, validation, surveillance, auditability and clear attribution of the final decision. In healthcare, an algorithm cannot be a "black box" that undermines the clinician-patient relationship; it must be a tool that makes it more informed.

Another challenge, also for Italy, concerns the completion of the data ecosystems on which AI is based.

Italy has developed the Fascicolo Sanitario Elettronico 2.0, the Health Data Ecosystem (EDS) and the European Health Data Space, "important steps towards greater data integration, but semantic interoperability between systems is still far from being fully achieved," says Bechini.

"Feeding the Fascicolo Sanitario Elettronico 2.0 with imaging data and diagnostic metadata means making available essential information for continuity of care, research and development of more effective AI models," explains Bechini.

"This aspect becomes even more relevant in the field of community medicine and the redesign of Diagnostic-Therapeutic-Assistance Pathways (PDTA), where the acquisition of diagnostic images in community healthcare settings is a fundamental element to ensure early diagnosis, timely management and real integration between hospital and local services".

"The true digital transformation, therefore, is not only about tools, but the ability to put data, images, processes and the patient at the center of a truly connected healthcare ecosystem," says Bechini.

The future of healthcare is not one in which medicine is governed by algorithms, but of a medicine assisted by data, regulated, validated and centred on the patient. Artificial intelligence can only really make a difference if it is integrated into clinical processes with accountability, scientific evidence and governance. Its value does not lie in making the diagnosis "automatic", but in making it more timely, more precise, more reproducible and more sustainable.

Alessandro Longo

is a journalist specializing in technology. He is Editor-in-Chief of AgendaDigitale.eu and a contributor to La Repubblica and Il Sole 24 Ore. In 2020, he published Artificial Intelligence: The Impact on Our Lives, Rights, and Freedoms (Mondadori Education, Milan).

BIBLIOGRAPHY

McKinney, S.M., Sieniek, M., Godbole, V. et al. International evaluation of an AI system for breast cancer screening. Nature, 577, 89-94, 2020. DOI: 10.1038/s41586-019-1799-6

Lång, K., Josefsson, V., Larsson, A.M. et al. Artificial intelligence-supported screen reading versus standard double reading in the Mammography Screening with Artificial Intelligence trial (MASAI): a clinical safety analysis of a randomised, controlled, non-inferiority, single-blinded, screening accuracy study. The Lancet Oncology, 24(8), 936-944, 2023. DOI: 10.1016/S1470-2045(23)00298-X 

Hernström, V., Josefsson, V., Sartor, H. et al. Screening performance and characteristics of breast cancer detected in the Mammography Screening with Artificial Intelligence trial (MASAI): a randomised, controlled, parallel-group, non-inferiority, single-blinded, screening accuracy study. The Lancet Digital Health, 7(3), e175-e183, 2025. DOI: 10.1016/S2589-7500(24)00267-X 

Gommers, J., Hernström, V., Josefsson, V. et al. Interval cancer, sensitivity, and specificity comparing AI-supported mammography screening with standard double reading without AI in the MASAI study: a randomised, controlled, non-inferiority, single-blinded, population-based, screening-accuracy trial. The Lancet, 407(10527), 505-514, 2026. DOI: 10.1016/S0140-6736(25)02464-X 

Ardila, D., Kiraly, A.P., Bharadwaj, S. et al. End-to-end lung cancer screening with three-dimensional deep learning on low-dose chest computed tomography. Nature Medicine, 25, 954-961, 2019. DOI: 10.1038/s41591-019-0447-x

Aerts, H.J.W.L., Velazquez, E.R., Leijenaar, R.T.H. et al. Decoding tumour phenotype by noninvasive imaging using a quantitative radiomics approach. Nature Communications, 5, 4006, 2014. DOI: 10.1038/ncomms5006

Coudray, N., Ocampo, P.S., Sakellaropoulos, T. et al. Classification and mutation prediction from non-small cell lung cancer histopathology images using deep learning. Nature Medicine, 24, 1559-1567, 2018. DOI: 10.1038/s41591-018-0177-5

Kather, J.N., Pearson, A.T., Halama, N. et al. Deep learning can predict microsatellite instability directly from histology in gastrointestinal cancer. Nature Medicine, 25, 1054-1056, 2019. DOI: 10.1038/s41591-019-0462-y

Fu, Y., Jung, A.W., Torne, R.V. et al. Pan-cancer computational histopathology reveals mutations, tumor composition and prognosis. Nature Cancer, 1, 800-810, 2020. DOI: 10.1038/s43018-020-0085-8

McGenity, C., Clarke, E.L., Jennings, C. et al. Artificial intelligence in digital pathology: a systematic review and meta-analysis of diagnostic test accuracy. npj Digital Medicine, 2024. DOI: 10.1038/s41746-024-01106-8

Oderda M. et al., Transperineal Free-Hand Prostate Fusion Biopsy with AI-Driven Auto-Contouring: First Results of a Prospective Study, Cancers, 2025, 17, 2381, DOI: 10.3390/cancers17142381.

Yang, A., Finkelstein, M., Koo, C., Doshi, A.H. Impact of Deep Learning Image Reconstruction Methods on MRI Throughput. Radiology: Artificial Intelligence, 6(3), 2024. DOI: 10.1148/ryai.230181

Gassenmaier, S., Afat, S., Nickel, M.D. et al. Deep learning-accelerated T2-weighted imaging of the prostate: Reduction of acquisition time and improvement of image quality. European Journal of Radiology, 137, 109600, 2021. DOI: 10.1016/j.ejrad.2021.109600 

Gassenmaier, S., Afat, S., Nickel, M.D. et al. Accelerated T2-Weighted TSE Imaging of the Prostate Using Deep Learning Image Reconstruction: A Prospective Comparison with Standard T2-Weighted TSE Imaging. Cancers, 13(14), 3593, 2021. DOI: 10.3390/cancers13143593