AI in medical diagnosis: what changes in practice for healthcare systems

The narrative that artificial intelligence will transform medicine in the future ignores what is already happening now. In the most image-intensive specialties — radiology, pathology, ophthalmology — tools with regulatory approval from the North American FDA already detect pulmonary nodules, diabetic retinopathy, stroke and suspicious lesions on mammography with clinical performance comparable to that of the specialist. IDx-DR, approved in 2018, diagnoses diabetic retinopathy without the need for in-person medical interpretation. Viz.ai detects large arterial occlusions in head CT scans and directly contacts the interventional neurology team. The bottleneck today is not the capacity of the tools — it is the ability of health systems to integrate them in an operationally coherent way.

What clinical evidence has already confirmed

The regulatory approvals curve in the area of medical imaging accelerated significantly between 2020 and 2025. The FDA cleared more than 700 AI and machine learning devices for clinical use by the middle of this decade, with radiology accounting for the absolute majority of these authorizations. Studies published in The Lancet, NEJM and Radiology demonstrate that breast cancer detection algorithms in mammography reduce false negatives and alleviate the burden of double reading — a mandatory practice in several European systems. In digital pathology, models trained to identify tumor subtypes in histological slides have already surpassed interobserver agreement among experienced pathologists in some specific categories of cancer.

This does not mean that AI replaces the radiologist or the pathologist. What the evidence consistently shows is that the human-algorithm combination performs better than either alone — especially in high-volume screening tasks, where fatigue and attention span introduce systematic errors. The most solid use case is not replacement: it is the amplification of diagnostic capacity without proportionally expanding the number of specialists.

What is happening in Brazil with regulation and adoption

In recent years, ANVISA has created a specific regulatory framework for software as medical devices — the so-called SaMDs (Software as a Medical Device), following the international classification ISO 13485 and aligned with the IMDRF guidelines. Diagnostic AI tools fall into this category and require registration or notification depending on the associated risk. The challenge in Brazil is not the absence of regulation: it is the speed of processing and the technical capacity of national manufacturers to prepare robust clinical dossiers.

Some Brazilian healthtech startups already have products approved or in an advanced notification phase with ANVISA — especially in the areas of chest X-rays for tuberculosis and pneumonia screening, and in fundus examinations. Hospital networks and health plan operators began to pilot them on an occasional basis, but the transition from pilot to systemic implementation still faces problems that are neither regulatory nor technical: they are infrastructure and process issues.

Integration with PACS, medical records and the real implementation problem

Any AI algorithm in medical imaging needs to connect to PACS — the image storage and communication system — and, ideally, the patient's electronic medical record to access the clinical context. This point of integration is where most projects stumble. The PACS installed in Brazilian hospitals vary greatly in version, supported DICOM protocol and openness to external APIs. Many legacy systems do not have standardized interfaces for receiving or exporting data in a way that modern AI tools can consume with low friction.

The practical result is that deployment that should take weeks often takes months — and not because of the AI itself, but because of the hospital's data architecture. Hospital IT teams are rarely familiar with required integration standards, and AI vendors often underestimate the effort to adapt to the customer's real environment. Projects that do not map this hidden integration cost before hiring tend to consume budget and political energy without generating measurable clinical results.

Civil liability and workflow design

The issue of legal responsibility in AI-assisted diagnosis is still far from being pacified in Brazil. The CFM (Federal Council of Medicine) has already stated that the final responsibility for the diagnostic act remains with the doctor — which is clinically sensible, but creates operational tension when the algorithm catches something that the radiologist has not prioritized or ruled out. This scenario is not hypothetical: it emerges naturally when AI is used as a screening tool, pre-selecting critical cases in reporting queues.

Workflow design needs to respond to this tension before it becomes contentious. How does the AI alert appear in the radiologist interface? Does the final report record that there was algorithmic assistance? Is there an audit trail that documents what the system flagged versus what the doctor concluded? These questions have no single answer, but they need to be answered before deployment — not after the first incident.

What health system leaders need to understand before piloting

Choosing an AI diagnostic tool based on the accuracy benchmark published in the paper is the equivalent of hiring an employee just based on their resume without checking references. Performance reported on validation datasets rarely replicates without degradation in a hospital production environment with older equipment, varying acquisition protocols, and a different population than that used in training. The selection process needs to include a local validation phase — preferably with retrospective data from the hospital itself — before any scale commitment.

Furthermore, the internal project sponsor matters as much as the technology. AI deployments in radiology that don't have a senior radiologist as an active advocate within the clinical team often die of disuse: the tool exists, but no one uses it systematically because it wasn't designed for the actual workflow. Clinical buy-in cannot be achieved with demonstration at a conference — it is achieved with the active participation of doctors in the configuration phase and with evidence generated from the hospital's own data.