In July 2026, a perspective article co-authored by scholars from Harvard, Dresden, Mayo Clinic, Tsinghua University, and other institutions—published in NEJM AI—presented three striking figures: among the 1,430 AI/machine learning medical devices cleared to date by the U.S. FDA, only five fall under the category of "modern pathology AI based on whole-slide images"; radiology-related devices number 1,094; and only about 4% of tissue slides in the U.S. are digitized for digital reading. Pathology—the gold standard for cancer diagnosis—remains one of the clinical domains with the lowest AI adoption rates. Paradoxically, precisely between 2025 and 2026, pathology AI delivered the highest-evidence-level clinical utility studies across the entire field of medical AI. The gap between evidence and adoption merits close examination.
## Layer One: Biomarkers That Predict "For Whom This Drug Works"
The first tier of the evidence chain appeared in the November 2025 issue of the Journal of Clinical Oncology. Researchers integrated digitized prostate biopsy histopathology images with clinical variables—including age, PSA level, Gleason score, and tumor stage—to train a multimodal artificial intelligence (MMAI) model. The model was trained on six NRG Oncology phase III randomized trials and validated on a seventh trial, RTOG 9202 (n = 1,192). In the validation cohort—with a median follow-up of 17.2 years—long-term hormonal therapy reduced the rate of distant metastasis from 26% to 17% (subdistribution hazard ratio [HR] 0.64, 95% CI 0.50–0.82). Crucially, a statistically significant interaction existed between the biomarker and treatment (P = 0.04): MMAI-positive patients (n = 785, 66%) derived substantial benefit from long-term therapy (HR 0.55, 95% CI 0.41–0.73), whereas MMAI-negative patients (n = 407) showed no benefit (HR 1.06, 95% CI 0.61–1.84, P = 0.84). The 15-year absolute risk difference for distant metastasis was 14 percentage points among MMAI-positive patients but zero among MMAI-negative patients. The authors termed this "the first validated predictive biomarker, to our knowledge, guiding duration of combined radiotherapy and hormonal therapy," implying that approximately one-third of high-risk patients could safely avoid an additional 24 months of drug exposure—and its associated toxicities.
External validation came from the STAMPEDE platform: in the July 2025 issue of The Lancet Digital Health, the locked algorithm was tested across four phase III trials within STAMPEDE, involving 112 centers and patients with advanced prostate cancer. Among 3,167 patients, each one-standard-deviation increase in MMAI score corresponded to a higher risk of prostate cancer–specific death (HR 1.40, 95% CI 1.30–1.51); when stratified by quartiles, risk separation persisted even within the same disease burden stratum—for non-metastatic, lymph-node–negative patients, the 5-year prostate cancer–specific mortality was 3% in the low-score group versus 11% in the high-score group. In October 2026, Annals of Oncology extended these findings further: in two sequential abiraterone phase III trials within STAMPEDE (n = 1,137 randomized), adding abiraterone to the ultra-high-risk MMAI subgroup (n = 268) improved 5-year metastasis-free survival from 62% to 81% (HR 0.47, 95% CI 0.31–0.70); the standard high-risk subgroup (n = 869) derived limited benefit (HR 0.83, 95% CI 0.63–1.09), with a statistically significant interaction (P = 0.02). Histopathological images thus began directly predicting "whether this drug works for this patient." Note, however, that this study was a post hoc analysis of randomized trial data; although the 75th-percentile threshold was prespecified, its retrospective nature still limits the strength of clinical recommendations.
Moreover, this is not a story confined to a single company. In February 2026, Clinical Cancer Research published an independent model—APIC—developed by Madabhushi’s team at Georgia Tech and Emory University. In the CHAARTED trial (metastatic population), APIC predicted survival benefit from docetaxel: APIC-positive patients (56.7% of cohort) had an overall survival HR of 0.52 (95% CI 0.31–0.85), while APIC-negative patients derived no benefit (HR 1.31, 95% CI 0.71–2.44), with a significant interaction (P = 0.022). Similarly, in NRG/RTOG 0521 (localized high-risk disease), APIC remained highly significant: APIC-positive patients (44.7% of cohort) had an HR of 0.49 (95% CI 0.26–0.92), with interaction P = 0.024; significance persisted after adjusting for clinical factors. The authors themselves listed limitations: retrospective design, need for prospective validation, and necessity for re-evaluation in the era of "triplet therapy."
## Layer Two: A Necessary Dose of Cold Water
In July 2026, a study published in the Journal of the National Cancer Institute asked another question: How does MMAI compare against "already-good tools"? Researchers applied the whole-slide biopsy–designed MMAI model to tissue microarrays (TMAs) from 424 radical prostatectomy specimens (414 successfully scored, 98%), with 10-year follow-up—a cohort with overall excellent prognosis: 10-year biochemical recurrence–free survival was 74%, and metastasis-free survival was 96%. In univariable analysis, MMAI was indeed prognostic: HR for biochemical recurrence was 1.04 (95% CI 1.02–1.06); HR for metastasis was 1.05 (95% CI 1.02–1.07). However, after adjustment for the CAPRA clinical scoring system, MMAI lost independent prognostic value; significance was retained only after adjustment for the cell cycle progression (CCP) genomic score. Correlations between MMAI and either CAPRA or CCP were weak (r < 0.35); the combined MMAI + CCP model achieved the highest discrimination for metastasis (C-index 0.76), comparable to CAPRA alone (0.75). The authors’ conclusion was restrained: MMAI does not outperform existing clinical tools but provides complementary information when integrated with genomic data—"warranting further refinement and validation in diverse clinical settings."
An easily overlooked detail here is applying a model designed for whole-slide biopsies to TMAs—an out-of-distribution test beyond its intended use. Yet the core message remains valid: demonstrating "treatment-benefit prediction" in trial cohorts is fundamentally different from proving "incremental value over a freely available clinical score" in routine care. The latter question remains unanswered.
## Layer Three: The Bottleneck Is Not the Algorithm
That NEJM AI perspective article ranked barriers to pathology AI adoption hierarchically: at the base lie infrastructure and economic challenges—costs of scanners, storage, and workflow integration; above them sit preanalytical variability post-digitization and clinician trust; next are immature implementation and governance pathways; and at the apex lies the equity gap in low- and middle-income countries. A 4% digitization rate means that deployment of most pathology AI is physically impossible: without digital slides, even the best algorithms have nowhere to run. This explains an apparent paradox—why the discipline with the highest-quality evidence has the lowest adoption rate. The article proposed two near-term translational pathways: using AI as a digital copilot to assist clinical assessment, and leveraging AI to empower clinical trial design.
An intriguing workaround appeared in Nature Communications in August 2026: a team from Memorial Sloan Kettering Cancer Center developed ALLocate—a low-cost "autonomous driving" add-on directly mounted onto conventional optical microscopes, designed for acute leukemia bone marrow smear screening. Training used >11,000 annotated regions and 130,000 annotated cells; validation employed 165 real bone marrow smears from multiple independent institutions: area under the curve (AUC) for region-of-interest detection exceeded 0.99; mean average precision at IoU threshold 0.50 (mAP@50) for cell detection reached 0.90; slide-level diagnostic accuracy was 88%—all achieved without requiring whole-slide scanners. While 88% slide-level accuracy remains far from clinical diagnostic standards (this judgment is ours, not the paper’s), its significance lies in the pathway: rather than waiting for pathology departments to complete costly full-scale digitization, bring AI directly to the microscope.
## My Three Judgments
The following synthesis emerges from reading all seven studies together—not conclusions drawn from any single paper.
First, clinical utility evidence for pathology AI truly advanced in 2025–2026: it moved beyond retrospective benchmarking against pathologist concordance to becoming a predictive biomarker rigorously validated via treatment–biomarker interaction tests across multiple phase III randomized trial cohorts. More importantly, two mutually independent teams produced isomorphic results across three distinct therapeutic decisions—duration of hormonal therapy, abiraterone addition, and docetaxel use. In terms of evidentiary density, this is rare across medical AI specialties.
Second, evidence generation and commercial interests are highly overlapping—making the next hurdle unmistakably clear. Within the MMAI evidence chain, the JCO abstract’s conflict-of-interest statement lists multiple Artera employees and their stock holdings; The Lancet Digital Health statement explicitly names two authors as company founders; both the Annals of Oncology and JNCI papers list multiple Artera employees among their author affiliations—but neither PubMed abstract includes a conflict-of-interest statement (full disclosure requires consulting the manuscripts); the lead investigator of the independent APIC model holds equity in three digital pathology imaging companies, disclosed in the abstract. These disclosures do not invalidate the evidence—they are, in fact, quite transparent—but they define what must come next: independent replication, and prospective interventional studies assigning treatment by biomarker status with patient-centered outcomes as endpoints.
Third, the adoption bottleneck is infrastructure—not algorithms. Single-digit digitization rates coexisting with trial-grade evidence confirm that the transmission belt from evidence to adoption breaks precisely at the digitization step. Thus, paths more immediately viable than "replacing pathologist slide review" include: AI as a copilot, AI-enabled clinical trial design, and low-cost hardware add-ons mounted directly onto microscopes.
## What You Can Take Away
If you are a prostate cancer patient or caregiver: duration of hormonal therapy, and decisions about adding abiraterone or docetaxel, now have candidate biomarkers for discussion—but until guideline adoption and prospective validation, they do not constitute standard practice. A key question to ask is: "Has this test undergone prospective validation?"
If you are a hospital administrator: the cost calculus of digital pathology is a prerequisite for AI adoption. Where you stall in the barrier hierarchy determines which AI solutions are relevant to you.
If you are a researcher: head-to-head comparisons against simple clinical tools should become the default requirement for publishing new models—the JNCI paper exemplifies exactly this.
If you are a general reader: the next time you see news about "AI predicting treatment benefit," ask two questions first—was it a post hoc analysis or a prospective trial? And against which comparator tool was it evaluated?
Who should set the rules for such tests entering clinical guidelines? Is your hospital’s pathology department digitized? We welcome your observations in the comments.
Sources: NEJM AI, online ahead of print, July 15, 2026 (PMID 42932556, https://doi.org/10.1056/AIp2600029); Journal of Clinical Oncology, November 10, 2025 (PMID 40239134); The Lancet Digital Health, July 2025 issue (PMID 40467357); Annals of Oncology, online ahead of print, June 6, 2026 (PMID 42250604); Clinical Cancer Research, February 4, 2026 (PMID 41324555); Journal of the National Cancer Institute, online ahead of print, July 10, 2026 (PMID 42429593); Nature Communications, August 27, 2026 (PMID 42805957).