A model trained on pathologists' pan-and-zoom logs caught every cancer-positive slide

Expect the 100 percent to travel furthest, and it is the figure that needs a denominator. The Live Science account does not state how many slides were in the test set ^[22]. Sensitivity is a fraction of the positive slides only, so on a set holding twenty positives, one miss would have printed as 95 percent ^[23]. The head-to-head with OpenAI's o3 carries more information, because both models worked the same labeled slides ^[14].

Read the false-alarm shares as precision instead. Of the slides Pathology-o3 called positive, 84.5 percent really were positive ^[18]. For o3, 46.7 percent were ^[19]. The share of positive calls that were wrong ran about 3.4 times higher for o3 ^[21], and o3 also missed 12.5 percent of the cancer-positive slides outright ^[20].

Strip it back and the design question is where to spend magnification. A whole slide can hold billions of pixels while the evidence of cancer occupies a tiny patch ^[3]. Splitting that slide into fixed-size patches treats every square of tissue as equally worth reading ^[1]. Pathology-o3 makes a low-resolution pass first, lets the behavior-trained model pick regions, and only then reads those at higher magnification ^[12]. Huang compared it to a search-and-rescue helicopter. "You don't start by inspecting one square meter of ground," he told Live Science ^[6].

I would most want to see the training data replicated. Most pathology systems learn from what a pathologist leaves behind at the end of a search, a labeled image or a diagnosis ^[24]. Huang's team recorded the search itself, and since raw logs include drift, overshoot and fiddling with magnification ^[8], the filtered output needed a check: they compared the regions their filter kept against eye-tracking data to confirm the software was capturing where the pathologists actually looked ^[9].

Pathologists stay in the loop on the language side as well. For each inspected region the model drafted a short rationale for why it was worth examining, and pathologists accepted, edited or rejected each draft; those judgements became further training data ^[10]. In one case the model called a region potentially metastatic and proposed zooming in to look for atypical cells ^[11].

The test also carries a structural limit. The slides were labeled by human pathologists before the test ^[14], so what was measured is agreement with a diagnosis already completed, not a change to one in progress.

The false positive share came out of a tuning choice. The researchers built the system to err toward flagging a region for another look instead of risking a miss, and Huang said that might explain the rate ^[17].