Two-word malaria counting: SAM 3 counts 28 infections, expert counts 29—where are the misses?
Maziyar Panahi fed blood smear photos from a Bangladeshi hospital into SAM 3, with only “cell” and “purple dot” as input; the model counted 28 infected red blood cells, while an expert counted 29. On a single L4, each word took 0.15 to 0.17 seconds. A closer look: 25 of the 28 matched the expert, 4 were missed, 3 were extra, and ring-stage parasites were not recognized. For research use, still far from clinical use.
Malaria microscopy still relies on the human eye. Under the microscope, a person counts whether there are parasites in red blood cells. Maziyar Panahi posted an experiment on X: blood smear photos from a hospital in Bangladesh were given to SAM 3, with only two words as input, “cell” and “purple dot.” The model counted 28 infected red blood cells; the expert counted 29.
The method is not complicated. SAM 3 comes from Meta, exported to ONNX via @wkentaro_, and runs on a single L4 in Hugging Face Spaces, taking 0.15 to 0.17 seconds per word. The decision rules were also spelled out: a red blood cell counts only if its outline appears completely in the field of view, and an infection counts only if a purple dot falls on a red blood cell. The smears come from the U.S. National Library of Medicine and are labeled for research use.
The numbers are very close. But breaking 28 and 29 apart makes the information more interesting. Of the model’s 28, 25 were the same cells marked by the expert. It missed 4 and added 3: 1 platelet and 2 red blood cells that the expert judged clean. Across 7 smears and 7 patients, it found 421 red blood cells versus the expert’s 445; it found 72 infected red blood cells versus the expert’s 112. It is especially poor at ring-stage parasites and recognizes dot-like parasites better.
This is the distance between a demo and a tool. Being off by one number makes it look almost there, but the structure of the errors shows it is not ready to replace microscopy. Which morphologies the false positives and misses fall on is more worth looking at than the overall score. The author himself wrote research use only. When someone asked where they could try it, he said he would add a harness layer for tool calling and precise measurement; the code is open source, and the app is hosted on Hugging Face.
Getting this far with two words is worth a look from people in the lab. But whether to use it first requires asking: on which kind of smear, at which parasite stage, and who will review it. Source: https://x.com/MaziyarPanahi/status/2116893021073523051
发布时间: 2026-11-02 04:42