Header: AMA CPT may soon face a crossroads for modernizing coding of digital pathology.
##
Update: The U.S. Senate wants to hear your concerns about the coding system and blocking innovation - here.
##
AMA CPT has had limited means of coping with digital pathology and AI, and in some ways, has even regressed. There is a single Category I CPT code for making 1 immunostained slide and interpreting it with computer assistance (88361).
For several years, AMA also made about 10 codes involving whole slide imaging in the PLA series, but labs that have tried to replicate this have told me that the rules have changed. If the biomarker requires immunohistochemistry, it may be judged to include "physician work" and therefore ineligible for PLA coding. (I worked with one lab whose service didn't involve any physician work, and was rejected anyway for "requiring physician work.") But if the service is based on conventional stains, like H&E, it may be judged as not having a bona-fide biomarker, and therefore ineligible for PLA. If some of the service is performed offsite, e.g. a cloud algorithm allowed by CAP distributed testing (here), or a separately prepared DICOM file, it may be judged as a multi-lab, not single-lab, service and therefore ineligible for PLA. If it involves "AI" (and NGS has already involved massive amounts of bioinformatics) it may get shunted into another coding system, a sort of registry for SaMD. It's like a game of Chutes and Ladders, where there are lots of chutes on every side and not a ladder in sight.
##
Meanwhile, outside of AMA CPT policy, the world is moving fast. Here are just four examples:
- Rakaee's report of deep learning immunotherapy guidance from digital imaging of H&E, reported in JAMA Oncology. Here.
- FDA authorizes a clinical trial biomarker (which may well become a companion diagnostic), used in NASH drug trials and based on software reading of trichrome and H&E sections. See FDA announcement, here. See the project's webpage for associated documents here. There, see particularly the FDA integrated review (PDF 52pp). See discussion at Linked In here.
- See similarly, Valanarasu et al., in Cell, on GigaTIME sophisticated proteomic reproduction from AI-assisted readings of H&E slides. Here. See insights from two authors at Linked In, here and here. Also Hayes.
- Meanwhile, and also based on FDA drug trials, see the novel "QCS" AI-assisted immunomarker reading, which can do things that humans simply can't do. It's a collaborative project for Roche, AstraZeneca, FDA, entry point here.
Is scientific reality is outrunning CPT structures?
OVERVIEW BY AI
Taken together, these four groups of articles point to a single, uncomfortable conclusion: the center of gravity in diagnostic pathology has shifted, but AMA CPT remains anchored to an older mental model of what a “test” is.
Across all four domains—ICI response prediction in lung cancer, FDA-qualified AI-assisted NASH scoring, population-scale inference of proteomic states from H&E, and AI-driven companion diagnostics such as AstraZeneca’s QCS—the core diagnostic value no longer resides in the stain, the slide, or even the individual pathologist’s interpretive act. It resides in computational extraction of clinically meaningful information from routine histology, often using H&E alone, sometimes executed in distributed or cloud environments, and frequently exceeding what unaided human vision can achieve.
The common thread is that biological insight is being unlocked without new wet-lab chemistry. These systems do not require novel antibodies, multiplex panels, or exotic reagents. Instead, they repurpose the most ubiquitous artifact in pathology—the H&E slide—into a high-dimensional data source. In several cases, including AIM-NASH and QCS, regulators have already accepted this transformation as sufficiently rigorous to support drug development, clinical trial endpoints, or companion diagnostics. The science, validation, and regulatory logic are already in place.
Against that backdrop, CPT policies that disallow coding because an algorithm runs in the cloud, or because an H&E-based method lacks a “bona fide biomarker,” are not neutral gatekeeping decisions. They function as structural blinders, preventing recognition of diagnostics whose value is computational rather than chemical. The result is a paradox in which FDA-qualified tools and drug-enabling diagnostics can be scientifically legitimate yet economically invisible under CPT.
If CPT continues to equate diagnostic legitimacy with stains, analytes, and on-site manual interpretation, it risks becoming misaligned with both modern pathology and modern therapeutics. The articles you cite are not edge cases or speculative futurism; they are signals that computational pathology is already doing clinically consequential work. A coding framework that cannot accommodate H&E-based or cloud-executed diagnostics is not being cautious—it is simply falling behind the science.
A. Deep Learning–Based Slide Reading for Immunotherapy Response in Lung Cancer
The Rakaee et al. JAMA Oncology paper represents a mature example of end-to-end predictive pathology: a deep learning model trained directly on routine H&E whole-slide images to predict response to immune checkpoint inhibitors (ICI) in advanced non-small cell lung cancer. Importantly, the model is not attempting to recapitulate a known biomarker such as PD-L1 or TMB. Instead, it learns higher-order spatial and morphological patterns—many of which are not explicitly nameable by human observers—that correlate with therapeutic response.
The work is notable for its external validation across multiple institutions and continents, which is essential given the notorious domain-shift problems in digital pathology. Performance is clinically meaningful: AUCs comparable to PD-L1 alone, and materially improved stratification when combined with PD-L1. Conceptually, this reframes the slide not as a passive substrate awaiting a stain, but as a rich, latent data object encoding immune contexture, tumor architecture, and stromal interactions.
From a regulatory and coding perspective, the study highlights a tension: the output is clinically actionable, yet it does not map cleanly onto existing CPT categories tied to stains, analytes, or discrete human interpretive acts. This is neither conventional histopathology nor a molecular assay—it is something genuinely new.
B. FDA-Qualified NASH Biomarker Based on Trichrome and H&E Slides
The FDA’s qualification of the AIM-NASH system marks a watershed moment: the first AI-based histologic drug development tool formally qualified under the Biomarker Qualification Program. Unlike many exploratory AI pathology efforts, AIM-NASH is tightly scoped to a specific regulatory context of use—assisting pathologists in grading NAS components and fibrosis in MASH clinical trials. Its goal is not replacement of human judgment, but reduction of variability that has long plagued trial enrollment and endpoint assessment.
Crucially, the system operates entirely on standard H&E and trichrome slides, leveraging AI to generate reproducible quantitative assessments where ordinal human scoring has proven noisy. The validation strategy is conservative by FDA standards, emphasizing non-inferiority to expert consensus reads, explicit rejection workflows, and continued human accountability. This makes the qualification less about algorithmic novelty and more about process discipline, statistical rigor, and transparency.
What is striking for the broader argument is that FDA has accepted an AI interpretation of conventional stains as a bona fide regulatory instrument, while CPT struggles to classify similar approaches even for payment. The biomarker’s legitimacy flows from clinical trial utility, not from molecular specificity or staining novelty—again underscoring that modern pathology value increasingly lies in computation rather than chemistry.
C. GigaTIME and Advanced Interpretation of H&E Slides
The GigaTIME framework pushes computational pathology into genuinely new epistemic territory. By training on tens of millions of paired H&E and multiplex immunofluorescence (mIF) cells, the model learns a cross-modal mapping that can infer spatial proteomic patterns directly from routine H&E slides. The result is a “virtual mIF” representation that enables population-scale analysis of the tumor immune microenvironment across tens of thousands of patients—something practically impossible with real mIF due to cost and scarcity.
Scientifically, this is less about prediction of a single outcome than about synthetic expansion of biological observability. GigaTIME enables discovery of combinatorial protein patterns, spatial entropy measures, and immune architectures that were previously inaccessible at scale. Yet its power also raises an important caution: the system does not measure proteins—it infers them. The outputs are probabilistic reconstructions grounded in learned correlations between morphology and molecular state.
This distinction matters enormously for downstream use. GigaTIME is transformative for hypothesis generation, stratification research, and systems-level oncology, but it demands epistemic humility. It exemplifies how AI can create new scientific objects—“virtual populations”—that are immensely useful, yet ontologically different from direct assays. Coding and regulatory frameworks are not yet designed to acknowledge this category at all. [See also Hayes' essay here.]
D. AstraZeneca’s QCS Platform and AI-Enabled Companion Diagnostics
AstraZeneca’s Quantitative Continuous Scoring (QCS) platform illustrates how computational pathology can succeed where traditional immunohistochemistry has failed. In the TROPION-Lung01 trial, conventional TROP2 IHC was not predictive of response to datopotamab deruxtecan. QCS, by contrast, quantified subcellular TROP2 distribution at single-cell resolution, distinguishing membrane versus cytoplasmic localization across entire tumor populations. This enabled definition of a predictive biomarker that materially improved progression-free survival stratification.
What is technically important is that QCS is fully supervised and biologically intentional. It is not a black-box correlate, but a system designed to measure a specific mechanistic hypothesis: that intracellular trafficking of TROP2, not absolute expression, drives ADC efficacy. This aligns the computational readout tightly with drug mechanism, which in turn supports regulatory acceptance and companion diagnostic development with Roche.
From a systems perspective, QCS shows how AI pathology becomes indispensable when biology exceeds the limits of human vision and ordinal scoring. Yet it also exposes CPT’s bind: the assay involves IHC, physician oversight, whole-slide imaging, and AI computation—precisely the combination that currently disqualifies it from multiple coding pathways simultaneously. Scientifically validated, FDA-relevant, and commercially essential—yet structurally homeless in CPT.
###
