How much does AI improve diagnostic accuracy in real-world pathology?
AI pathology models consistently boost diagnostic accuracy when used alongside pathologists. In a randomized crossover reader study for gastric cancer, pathologists using the AI model GRACE improved their diagnostic accuracy from 82.0% to 89.9%, and the odds of making a correct diagnosis nearly doubled (odds ratio 1.987) [1]. The same study also found that AI assistance reduced diagnostic time by 14.9% and increased diagnostic confidence by 9.0% [1].
For prostate cancer, an AI system called Paige Prostate achieved 99% sensitivity and 100% negative predictive value at the patient level, meaning it rarely missed a cancer and could reliably rule out disease [5]. In that study, the AI actually identified four additional patients with cancer that three experienced pathologists had initially classified as benign or suspicious but not malignant [5]. This shows AI can catch cancers that even specialists miss.
In breast cancer, both pathologist-based and AI-based assessments of tumor-infiltrating lymphocytes (immune cells in the tumor) were equally predictive of whether patients would respond to chemotherapy [4]. However, in patients with low lymphocyte levels (non-LPBC), the AI method was better at distinguishing who would benefit from treatment (odds ratio 4.1 for achieving a complete response) [4]. This suggests AI can add value where human assessment is less discriminating.
Can AI save pathologists time and reduce their workload?
Yes, AI can dramatically reduce the time pathologists spend reviewing slides. The Paige Prostate AI system for prostate cancer provided an estimated 65.5% reduction in diagnostic time for the material analyzed [5]. That means for every hour a pathologist would normally spend, the AI could cut it to about 21 minutes.
The gastric cancer AI model GRACE also streamlined workflows. Under strict safety criteria (requiring 100% certainty to rule out or rule in cancer), the AI could still triage up to 69.6% of malignancy-diagnosis cases and 46.8% of MMR-IHC (a type of molecular test) follow-up requests, meaning those cases could be handled without full pathologist review [1]. For precancerous lesions, the AI-assisted workflow could triage 60.7% of atrophy cases and 82.7% of intestinal metaplasia cases while maintaining non-inferior performance to senior pathologists [1].
In a digital collaborative care model for irritable bowel syndrome (IBS), AI algorithms trained on a large gastrointestinal dataset helped personalize care plans, and 86% of participants experienced clinically significant symptom relief (≥50-point reduction on the IBS symptom severity scale) [3]. While this is not a pathology model per se, it shows how AI can support clinical decision-making and reduce the burden on healthcare systems.
Does AI help with prognosis and treatment decisions?
AI pathology models can provide prognostic and predictive information that directly guides treatment. In gastric cancer, the GRACE model achieved strong performance for molecular profiling (macro-AUC 0.8682) and prognostic prediction, meaning it could help determine which patients are likely to have aggressive disease and might benefit from more intensive therapy [1].
For non-alcoholic steatohepatitis (NASH), AI-based digital pathology detected treatment-induced fibrosis regression that conventional scoring systems missed entirely [2]. Patients who were judged as 'unchanged' by standard scoring actually showed measurable reductions in fibrosis when analyzed by AI, particularly in the perisinusoidal regions around areas of fat reduction [2]. This gives clinicians a more sensitive tool to assess whether a treatment is working.
In lymphoid neoplasms (blood cancers), next-generation sequencing panels—which are often analyzed with AI tools—provided prognostic value in 59.4% of cases and predictive value for therapy response or resistance in 32.6% of cases [6]. This means AI can help oncologists choose the right drug for the right patient.
A multimodal AI assistant called PathChat, which can answer open-ended questions about pathology images, produced responses that pathologists preferred over those from general-purpose AI like GPT-4V [7]. While not yet deployed in routine clinical care, such tools could support education, research, and human-in-the-loop decision-making.
About These Sources
This answer is built on 7 peer-reviewed studies — published from 2021 to 2026, 5 from 2024 or later, 2 in Q1 journals, collectively cited 541 times — selected as the most relevant from 9 studies that passed quality screening, drawn from 79 papers retrieved from a database of over 500 million.
Sources used in this answer
A Pathology Foundation Model for Gastric Cancer with Real-World Validation
In a randomized crossover reader study, the gastric cancer AI model GRACE improved pathologist diagnostic accuracy from 82.0% to 89.9%, nearly doubled the odds of a correct diagnosis (OR 1.987), reduced diagnostic time by 14.9%, and could triage up to 69.6% of malignancy cases under strict safety criteria.
Digital pathology with artificial intelligence analyses provides greater insights into treatment-induced fibrosis regression in NASH
In a study of 99 NASH patients from a clinical trial, AI-based digital pathology detected treatment-induced fibrosis regression that conventional scoring missed, revealing that fibrosis reduction begins in perisinusoidal regions around areas of steatosis reduction.
First Real-World Evidence of an AI-Enhanced Digital Collaborative Care Model to Improve IBS Symptoms.
In a prospective single-arm study of 202 IBS patients, an AI-enhanced digital collaborative care model produced a mean 140-point reduction in IBS symptom severity, with 86% of participants achieving a clinically significant ≥50-point reduction.
Abstract PS2-08-26: Pathologist- and artificial intelligence-based TILs assessment in patients with early triple-negative breast cancer treated with neoadjuvant chemo-immunotherapy: real-world evidence from a nationwide cohort
In a nationwide cohort of 337 early triple-negative breast cancer patients, both pathologist-based and AI-based tumor-infiltrating lymphocyte (TIL) assessments were predictive of complete response to chemo-immunotherapy; AI-TILs provided better discrimination in patients with low lymphocyte levels (OR 4.1 for pCR).
Independent real‐world application of a clinical‐grade automated prostate cancer detection system
In an independent real-world evaluation of 600 prostate needle core biopsies from 100 patients, the AI system Paige Prostate achieved 99% sensitivity and 100% negative predictive value at the patient level, identified four additional cancer cases missed by three pathologists, and reduced diagnostic time by 65.5%.
Next-generation sequencing for lymphoid neoplasms: Real-world utility from a clinical assay.
In a retrospective review of 384 lymphoma cases, next-generation sequencing panels provided prognostic value in 59.4% of cases and predictive value for therapy response or resistance in 32.6% of cases, aiding clinical management.
A multimodal generative AI copilot for human pathology
PathChat, a multimodal vision-language AI assistant for pathology, achieved state-of-the-art performance on multiple-choice diagnostic questions and produced responses that pathologists preferred over those from GPT-4V in open-ended queries.