The development of multimodal AI, which combines data from multiple sources, including pathology images and genomics, and data reported as text from clinical reports, is another substantial advance in computational pathology50,61,62,63,64,65. Multimodal approaches facilitate a more comprehensive understanding of disease mechanisms and patient outcomes. For example, Spratt et al.61 developed a multimodal AI approach using pretreatment prostate tissue images and clinical data (at a median follow-up duration of 10 years) from 5,727 patients enrolled in five phase III trials evaluating radiotherapy with or without androgen-deprivation therapy (ADT). The model predicted distant metastases and identified patients likely to benefit from ADT. Validation using data from the NRG Oncology/RTOG 9408 trial, with a median follow-up of 14.9 years, showed that ADT significantly reduced the occurrence of metastases in 34% of patients classified as model-positive, while providing no benefit to the 66% classified as model-negative61. This study highlights the predictive value of AI and its potential role in improving clinical decision-making.
Since 2019, the number of FDA approvals for devices based on AI and/or machine learning has increased substantially, rising from fewer than 70 to currently >1,000 (refs. ). This growth has prompted the evolution of regulatory frameworks, with the FDA and the European Union (EU) Conformité Européenne (CE) adapting guidelines to regulate AI-enabled medical devices. Notably, the Artificial Intelligence Act proposed by the EU in 2024 introduces a comprehensive legal framework, establishing EU-wide rules on data quality, transparency, human oversight and accountability through a risk-based approach. In parallel to developing this regulatory framework, AI-based tools such as Concentriq AP-Dx and Galen Prostate obtained CE marking (under the EU's In-vitro Diagnostic Medical Devices Regulation (IVDR)).
In the USA, the FDA authorized a De Novo Classification Request for the AI-based prostate cancer detection tool Paige Prostate in 2021, followed by 510(k) clearance decisions for other AI-based pathology tools developed for diagnostic purposes such as PathAI, Roche Digital Pathology Dx and Concentriq AP-Dx.
Nevertheless, approvals of AI-based tools for digital pathology are strikingly low relative to those for other medical applications. As of December 2024 (ref. ), radiology dominates the software-as-a-medical-device space in the USA by a substantial margin (76% of all FDA-approved devices), followed by cardiovascular disease (10%) and neurology (4%), while only three devices (<0.5%) have been approved for pathology (Fig. 1). This trend reflects a slower than anticipated integration of AI in digital pathology, especially considering that approvals of radiology devices typically have shorter median approval wait times (126 days) than those in other specialties (including 193 days for pathology).
In the USA, Current Procedural Terminology (CPT) codes provide a standardized framework for health-care professionals to accurately report the use of medical services and procedures, ensuring consistency and efficiency in claim processing, medical care review and insurance reimbursement. Interestingly, a wide range of CPT codes used for radiological procedures cover imaging services (codes 70010-76499). Meanwhile, digital pathology remains in the nascent stage of integration into clinical practice. Before 2019, pathology-related CPT codes primarily covered conventional diagnostic services, such as tissue biopsies and microscopic examinations, without specific provisions for digital or computational pathology. In 2024, 30 new Category III add-on codes, which are temporary CPT codes used to track new technologies, were added to the previously existing 13 codes. These new codes were introduced for services involving digitization of glass slides, which currently have no established payment rates, and they are expected to become established CPT codes (of Category I).
In the EU, some public reimbursement systems support digital health initiatives with funding from various sources to establish digital pathology pipelines. For example, in Germany fast-tracked reimbursement pathways for digital health applications are progressing through the Digital Health Applications (DiGA) programme. This programme provides a fast-track process for regulatory approval and reimbursement, allowing clinicians to prescribe approved digital health applications that are fully reimbursable by statutory health insurance, covering a large population. In France, through the PECAN programme, similar avenues are being explored for digital health solutions, including telemonitoring and digital therapeutics. Many European nations, such as the UK, Netherlands and Belgium, also reimburse AI-based in vitro diagnostic tests with funds from their national health-care reimbursement schemes.
In Asia, Japan and South Korea have established regulatory pathways for AI-based or machine learning-based medical devices, emphasizing safety, transparency, and performance. Other nations, including China and India, have introduced preliminary guidelines aimed at addressing challenges associated with AI technologies within their health-care systems.
Reimbursement frameworks are evolving to accommodate emerging technologies in digital pathology. Among these, LDTs and in-house devices are having a crucial role in addressing specific diagnostic needs within individual laboratories. Originally, LDTs were low-risk, specialized assays developed in the USA, created for specific, local patient populations and used in small volumes. These tests were tailored to address unique clinical needs, and thus they were exempt from rigorous FDA oversight and operated under the Clinical Laboratory Improvement Amendments (CLIA) certification framework. However, technological advances and shifts in business practices, combined with the ability to rapidly transport patient samples nationwide, have led to a broader use of LDTs. These tests are now applied to a larger, more diverse population, with major laboratories processing specimens from a wide geographical area. Their use is covered by a type of CPT code referred to as Proprietary Laboratory Analyses (PLA) codes. For example, the ArteraAI Prostate Test is an LDT available through a CLIA-certified laboratory designed to help clinicians to identify patients with localized prostate cancer who might benefit from ADT. This test has the unique PLA code 0376U. The inclusion of the ArteraAI Prostate Test in the US National Comprehensive Cancer Network (NCCN) guidelines for prostate cancer management from 2024 with a category 2A recommendation (indicating uniform consensus based on low-level evidence) emphasizes the growing relevance of AI-enabled tools for personalized treatment. Yet, concerns about the reliability of LDTs persist. In 2015, the FDA published 20 case studies detailing instances in which LDTs produced inaccurate results that led to inappropriate or delayed treatments, particularly in conditions such as cancer and heart disease. In response to these concerns, the FDA proposed new regulations in 2023 for LDTs to have the same stringent safety and efficacy standards as other medical tests, aiming to enhance patient protection and ensuring the reliability of LDTs. However, some experts have argued that this shift could lead to delays in testing for rare diseases and in responding to future health crises because laboratories developing tests with this purpose would first need to navigate the lengthy and costly FDA approval process.
The EU's IVDR framework for in-house devices is similar to the FDA's approach to LDT regulation. The IVDR framework allows health-care institutions to develop and use in-house devices internally under specific conditions, ensuring safety and reliability. The main difference between the EU and US approaches is that Article 5 of the IVDR imposes more stringent requirements on in-house devices. These requirements include proving that no CE-marked equivalent exists or that the in-house device addresses specific patient needs that CE-marked in vitro diagnostic tests cannot meet. Additionally, in-house devices must be used only within the owner institution and cannot be transferred to other entities. This approach enables European laboratories to continue innovative diagnostic practices while gradually harmonizing with regulatory expectations for safety and performance.
