What a General-Purpose AI Got Right (and Wrong) on 149 Expert-Graded Morphology Images
How well does a general-purpose AI identify cells? I tested one on 149 expert-graded morphology images. Here is what it got right, where it struggled, and why a human still needs to review it.
Draft. Only visible on preview links, not on the live site.
Like many people in our field, I feel a mix of excitement and fear about where AI is heading. I use AI tools heavily in my own work, and that got me wondering how well the current top models actually perform on something we do every day: identifying cells under the microscope.
So I pulled 149 expert-graded morphology images, the kind labs use for education and competency, covering peripheral blood, body fluids, urine sediment, and synovial crystals. I gave them to Claude Opus 5.5, one of the most capable AI models available right now. I did not give it a word bank or any example images. It saw only the image and a request to identify the target.
It matched the reference answer on 135 of the 149 images, or 90.6%.
This was a personal experiment, not a research study. It was one model in one session on one set of images, so your results may vary, and the next version of the model will likely perform differently. Here is what I found.
The Setup
Opus 5.5 is a general-purpose model that was not built or tuned for cell morphology. The set included some advanced educational challenge images. Every answer was compared against the reference answer key. There was no partial credit, and if the correct answer only showed up as a differential, it was scored as incorrect.
The original pool had 196 images, but many repeat from year to year. I removed 45 repeats and excluded two images that were not suitable for grading, which left 149.
One caveat is worth stating. Images like these circulate online, and I cannot rule out that the model saw some of them during training.
The Results
Standard images scored 125 of 137 (91.2%), and the educational challenge images scored 10 of 12 (83.3%). The model correctly identified a wide range of cells and elements:
- RBC morphology: schistocytes, sickle cells, target cells, teardrops, acanthocytes, echinocytes, stomatocytes, macro-ovalocytes, nRBCs, Howell-Jolly bodies, Pappenheimer bodies
- White cells: all normal types, Pelger-Huët, reactive and large granular lymphocytes, toxic changes, Döhle bodies, blasts with Auer rods, faggot cells
- Malignant cells: lymphoma cells including Burkitt, metastatic carcinoma clusters, mitotic figures
- Organisms: malaria, intracellular bacteria, yeast, Cryptococcus, Trichomonas, clue cells
- Body fluids: siderophages, erythrophages, alveolar macrophages, mesothelial cells, bronchial lining cells
- Urine and synovial: MSU, CPPD, cholesterol, uric acid, bilirubin, leucine, triple phosphate, ammonium biurate; RBC, WBC, granular, and waxy casts
For context, published evaluations of general-purpose models have mostly been discouraging. A 2024 evaluation of GPT-4o reached 78% accuracy on 100 white cell images, and that was with example images included in the prompt. A study in the Journal of Applied Laboratory Medicine found that ChatGPT-4 did not detect any of the blood parasites on peripheral smears. A 2025 preprint found that purpose-built hematology models outperformed general vision-language models. None of these are direct comparisons, since the images, tasks, and scoring all differ, but the gap suggests the models are improving quickly.
Where It Struggled
Here is every miss, grouped by specimen type, along with the confidence the model gave each answer.
| Specimen | Model’s answer | Reference answer | Confidence |
|---|---|---|---|
| Blood | Abnormal promyelocyte (APL) | Metamyelocyte | High |
| Blood | Abnormal promyelocyte (APL) | Myelocyte | High |
| Blood | Promyelocyte | Myelocyte | Moderate |
| Blood | Micromegakaryocyte | Promyelocyte | Low |
| Blood | Plasma cells | Blast | Low |
| Blood | Reactive lymphocyte | Plasma cell | Moderate |
| Blood | Intracellular ring form (parasite) | Platelet overlying an RBC | Moderate |
| Body fluid | Reactive mesothelial cluster | Malignant cells, non-hematopoietic | Moderate |
| Body fluid | Reactive mesothelial cluster | Malignant cells, non-hematopoietic | Moderate |
| Body fluid | Vacuolated macrophage | Mott cell | Moderate |
| Urine | Calcium oxalate monohydrate crystals | RBCs | Moderate |
| Urine | Sulfonamide crystals | Calcium oxalate crystals | Moderate |
| Urine | Neutrophil | Renal epithelial cell | Moderate |
| Urine | Renal tubular epithelial cell cast | RBC cast | Moderate |
Myeloid maturation was the clearest weak spot in blood. Four misses involved promyelocytes, myelocytes, and metamyelocytes, a distinction that depends on subtle changes in granules, chromatin, and nuclear shape. Two of those misses were the most concerning errors in the set, and I cover them below.
The body fluid misses were interesting. Twice, the model called a cluster of malignant cells a reactive mesothelial cluster. Telling reactive mesothelial cells from malignant cells is one of the hardest calls in body fluid morphology, and in most labs a cluster like that would go to pathology for review either way. The model rated both answers Moderate confidence, which tells me it was not sure either.
Urine was the weakest specimen type, accounting for 4 of the 14 misses, all at Moderate confidence. I think that comes down to how urine is examined. Blood smears are stained and spread flat, but urine sediment is unstained and viewed in a wet prep under a coverslip, where cells and casts sit at different depths and angles. A red cell turned on its edge can look a lot like the oval or dumbbell form of calcium oxalate monohydrate, and the model called those red cells crystals even though it listed red cells in its differential. The renal epithelial cell called a neutrophil and the RBC cast called a renal tubular epithelial cast fit the same pattern, since there is no stain to separate cells of similar size. My guess is that a model trained specifically on urine sediment images would close much of this gap, though even purpose-built urine analyzers still struggle with some of these calls.
A few other misses are classic traps that show up on graded images precisely because people miss them. These include a platelet sitting on top of a red cell that looks like an intracellular parasite, a Mott cell whose Russell bodies resemble vacuoles, and a plasma cell read as a reactive lymphocyte. In four of the 14 misses, the correct answer appeared in the model’s differential but was not its first choice.
What the Confidence Scores Showed
For every answer, the model also rated its own confidence. This turned out to be the most useful finding.
| Confidence | Images | Correct |
|---|---|---|
| High | 112 | 98.2% |
| Moderate | 34 | 70.6% |
| Low | 3 | 1 of 3 |
When the model rated an answer High, it was right 110 out of 112 times. Moderate answers were wrong about three times in ten, and there were too few Low answers to draw anything from. If a reviewer had checked only the Low and Moderate answers, 37 of the 149 images, they would have caught 12 of the 14 errors. Accuracy after that review would have been 147 of 149, or 98.7%.
The two errors that review would have missed were the two High-confidence myeloid calls. The model looked at a metamyelocyte and a myelocyte and called both abnormal promyelocytes consistent with APL. In a real lab, that would have been a false alarm on a hematologic emergency. An overcall is safer than a miss, but it is a good reminder that high confidence cannot mean no review.
Final Thoughts
I will be honest: I was shocked by these results. A model that was never designed for morphology, with no word bank and no examples, matched the expert answer on roughly nine out of ten images.
Will AI eventually replace some of the work we do at the microscope? Maybe. I do not think anyone can answer that honestly yet. What I can say is that these results remind me more of autoverification than replacement. Most of us already trust rules to release normal chemistry results while a technologist reviews the exceptions. Autoverification did not replace the chemistry tech. It changed where the tech spends their time.
I expect AI in morphology, and probably in other parts of the lab, to follow the same pattern. The system handles what it is confident about, a human reviews everything else, and someone keeps auditing the confident answers too. Those two APL overcalls are exactly why that last part matters.
So for now, I am more curious than afraid. I plan to keep running experiments like this as the models improve, and I will share what I find. If you are testing AI in your own lab, I would love to hear how it is going.