A whole-slide image at 40x magnification is roughly 100,000 × 100,000 pixels. At 20x it is still around 50,000 × 50,000. There is no monitor large enough to display one at native resolution, no human eye that can scan every tile systematically without losing context, and no reading session long enough to match the attention the image deserves when a pathologist is looking at case number 30 of the day. The question of where pre-reading adds value in digital pathology is really a question about how humans navigate that scale problem — and where they fail.
We have spent considerable time thinking about this while building Histolyx's pathology module. What follows is not a product pitch. It is an honest account of where tile-level pre-reading genuinely shifts how pathologists work, and where the current state of the technology still requires you to think carefully before relying on it.
The Scale Problem Is Not What You Think
The conventional framing is that WSIs are too large for humans to review exhaustively, so AI helps by reducing what the pathologist needs to look at. That framing is correct but incomplete. The harder problem is not the total pixel count — it is the combination of scale with non-uniform lesion distribution and the need to maintain context across zoom levels simultaneously.
When a pathologist reviews a tissue section at low power (2x or 4x), they are building a spatial map: the overall architecture, the relationship between structures, the areas that look architecturally abnormal. They then zoom to 10x or 20x to resolve ambiguous regions. This mental map is reconstructed from memory as the viewer navigates — and it degrades quickly when a case is complex or when the reading session runs long.
A tile-level classifier working at, say, 20x extracts feature vectors from non-overlapping tiles, then aggregates them into a slide-level or region-level prediction. The output of interest to us is not the slide-level label ("positive / negative") — that is a binary verdict that replaces the pathologist, which is not what we are building. The output we care about is a ranked spatial map: which tiles are most likely to contain diagnostically relevant structure? That is what a pre-reader should surface.
Region-of-Interest Surfacing vs. Classification
This distinction matters enormously in practice. A pre-reading system that outputs a heatmap saying "look here first, here second, here third" changes how a pathologist navigates a slide. A system that outputs a binary "suspicious / not suspicious" verdict short-circuits clinical judgment and creates documentation problems.
In Histolyx's pathology workflow, the output is an attention overlay: a tile-level confidence map that highlights regions above a configurable threshold, surfaced as a navigational aid inside the viewer. The pathologist opens the case and sees which regions the model flagged — they can navigate directly to those tiles, or follow their own protocol and use the overlay as a sanity check. The model is never presented as having read the case. It is presented as having pre-screened for regions worth closer inspection.
Consider how this plays out with a prostate core biopsy batch. A pathologist at a mid-size outpatient lab might receive 40–60 cores in a session. Each core is a separate slide. In a conventional workflow, every slide gets opened, panned, and reviewed at multiple magnifications. With Histolyx's overlay, the pathologist can open each slide and immediately see whether the pre-reader flagged any region. Slides with no flagged regions are still reviewed — we are not suggesting otherwise — but the pathologist can move through those cases with appropriate speed. Slides with flagged regions get more systematic attention first.
We are not saying that un-flagged slides are always clean. False negatives remain a real concern in any pre-reading system, and pathologists should calibrate their reliance accordingly. What the overlay changes is the navigation efficiency for high-volume sessions.
What Tile-Level Analysis Actually Captures
The features a tile-level model encodes at 20x are fundamentally morphological: nuclear size, nuclear-to-cytoplasmic ratio, mitotic figure density, gland architecture, stromal changes. These are the same features a pathologist uses at that magnification. What the model lacks is multi-scale context — it cannot simultaneously hold a tile-level nuclear feature and a slide-level architectural feature in a single representation the way a human reviewer can by scrolling.
This is why Histolyx's pathology model works at two scales: a 20x tile encoder for cellular morphology, and a 5x context encoder that captures tissue architecture over a larger field. The attention weights from both scales are combined to produce the heatmap. In internal testing on H&E slides, this two-scale approach reduced the number of architecturally significant but morphologically subtle regions that the model failed to flag — cases where individual tiles looked unremarkable but the glandular pattern at lower power was abnormal.
We are still working on multi-scale coherence for borderline cases, and the current model is strongest on adenocarcinoma patterns where both scales carry clear signal. Rare histological subtypes and tumors with significant necrosis remain harder — not because the model misses obvious findings, but because the confidence scores for those regions are less reliable as ranking signals.
Where This Fits in a Digital Pathology Lab Workflow
Pre-reading for digital pathology only makes sense if the lab has already moved to a digital workflow. We are not in the business of advising on scanner procurement, but the practical dependencies are worth naming: a whole-slide scanner, a viewing platform that accepts overlay data, and a LIS integration or worklist management system that can deliver cases to Histolyx and receive heatmap annotations back.
The integration path we use is straightforward: the scanner or image management system pushes the WSI file to Histolyx via a configured endpoint. Histolyx processes it asynchronously — typical turnaround for a single H&E slide at 40x is under four minutes on current infrastructure — and writes the annotation data back in a format the viewer can render. For labs using platforms with native overlay support, this requires no additional software on the pathologist's workstation.
The more nuanced operational question is how to handle the annotation data in the LIS or reporting workflow. Pathologists need to document what they reviewed, and the presence of an AI overlay should not create ambiguity about whether the pathologist reviewed the full slide or only the flagged regions. Our recommendation is to treat the heatmap as a navigation aid with no special status in the report — similar to how a grossing note guides initial attention without replacing microscopic review.
Calibration: What Thresholds Actually Mean
The confidence threshold for flagging tiles is configurable, and the choice matters. Setting it low means more tiles are highlighted — less likely to miss something, more likely to create noise that the pathologist ignores. Setting it high means the overlay is conservative and high-fidelity, but edge cases fall out.
We default to a threshold tuned for recall over precision — we would rather surface a region the pathologist dismisses after a 10-second look than have them miss a subtle finding because the model was not confident enough to flag it. That said, labs with very high slide volume and well-established protocols sometimes prefer a higher threshold to keep the overlay focused. We expose that setting per-lab, and we track how often flagged regions correlate with pathologist annotations in the final report to help labs understand what their chosen threshold is actually catching.
One finding from our calibration work: the threshold that produces optimal ranking performance on aggregate is almost never the same as the threshold pathologists in a given lab actually prefer. Pathologists develop an intuition for the overlay over time and want to tune it toward their individual cognitive style. This is not a problem — it reflects legitimate expert judgment — but it means a pre-reading system that offers only a global threshold is incomplete.
Honest Limits of the Current Approach
Tile-level pre-reading for WSI is further along for H&E histology than for IHC-stained slides, and further along for carcinoma patterns than for mesenchymal tumors, lymphoma subtypes, or inflammatory conditions. The model was trained on a mix of public and internal data with strong representation of common adenocarcinoma patterns — not because we chose to specialize narrowly, but because that is where training data density is highest.
Rare tumor types are a genuine gap. If a lab reads a high proportion of soft-tissue sarcomas, neuroendocrine tumors, or pediatric tumors, the current pathology module is not calibrated for those populations and should be evaluated carefully before use. We are not suggesting it will produce systematically wrong results — we are saying the confidence scores are less reliable as a ranking signal for those cases.
The other honest limit is that pre-reading addresses navigation efficiency, not diagnostic accuracy. If a finding is genuinely difficult — a borderline atypical proliferation, a pattern that requires clinical context to interpret, a case where reasonable pathologists disagree — a pre-reading heatmap does not resolve that difficulty. It gets the pathologist to the right tile faster. What happens next is still expert judgment.
For labs evaluating whether to deploy pre-reading for pathology, the question worth asking is not "will this catch findings I would otherwise miss?" That framing creates unrealistic expectations and puts AI in a role it should not occupy in current practice. The better question is: "will this change how my pathologists spend their attention in a high-volume session?" For labs doing digital pathology at scale, the answer is yes — in ways that are measurable and operationally meaningful.