The FDA's list of authorized AI-enabled medical devices passed 950 entries in 2024. Roughly three quarters of it is radiology. Pathology, a specialty whose entire practice is visual pattern recognition on images, contributes almost nothing.
That looks like a puzzle until you ask a duller question: were the images ever digital?
Radiology solved its data problem in the 1990s
Radiology digitized for reasons that had nothing to do with machine learning. Film was expensive to store, slow to retrieve, and impossible to be in two places at once. DICOM standardized the format and the transmission protocol, and PACS systems became standard infrastructure.
By the time deep learning arrived, radiology had three decades of standardized, portable, machine-readable images with structured reports attached. It handed the field a training corpus nobody had built for that purpose.
Pathology did not
Most pathology laboratories still cut glass slides and read them through a microscope. That workflow is cheap, reliable, and produces no digital image at all. A model cannot analyze an image that was never created.
Whole slide imaging had to be invented, validated, and authorized for primary diagnosis before any of this became possible. That authorization is recent, and the scanners are expensive, and a single slide can produce a file measured in gigabytes.
The economics are the actual barrier
Here is the part that explains why adoption has been slow even after the technology existed. Digitizing a pathology laboratory costs real money, and on its own it mostly reproduces what the microscope already did. Remote reading and easier consultation are genuine benefits. They are not obviously worth the capital cost.
Radiology never faced this problem, because PACS paid for itself in film and storage costs before anyone mentioned AI. Pathology has to justify digitization largely on the value of a computational layer that does not exist yet, which is a much harder internal sell.
What is actually authorized
The landmark is a single adjunct device that flags regions suspicious for prostate cancer on digitized biopsy slides, authorized through De Novo in 2021 as the first FDA-authorized AI application in digital pathology. It is an adjunct for one tissue type and one cancer, it requires a validated whole slide imaging pipeline as a prerequisite, and the pathologist remains the diagnostician.
The research literature is far ahead of that. Published work covers grading, biomarker prediction from morphology alone, prognostic modeling, and mutation inference from histology. Almost none of it is authorized, and the distance between "published in a good journal" and "cleared for clinical use" is the widest in this specialty.
The pattern generalizes
Specialties acquire AI in the order their data becomes machine-readable, not in the order they need it.
Cardiology has authorizations because ECG and echocardiography produce standardized digital signal. Ophthalmology has them because retinal photography is cheap and standardized with a clean diagnostic question. Psychiatry has almost none, because the data is conversation and observation with no standard digital form and no clean reference standard.
That makes digital infrastructure the leading indicator worth tracking. Where whole slide imaging adoption goes, pathology AI follows several years later, and the sequence is not going to run in the other direction.
See the current pathology device records and the pathology specialty hub for where things stand now.