Automated Leukocyte Classification

The problem

The laboratories already use a digital system that photographs the cells and proposes a pre-classification. It does not remove the review step. Someone still has to look at every pre-classified image, catch the errors, and confirm the final distribution — and on a busy technical platform that can occupy roughly four biologists and specialist technicians at once.

The work demands sustained visual concentration. Reviewing large numbers of images invites attentional fatigue, and fatigue lets incorrect classifications survive. Working through the data, I found mislabelled and inconsistent entries among the existing labels — some from genuine morphological ambiguity, some from variable image quality, some from human error at initial validation. That noise has to be found and cleaned before it reaches a training set.

There is a second, structural problem. The commercial model is a black box: its performance, its cell categories and its capacity to adapt are all determined by the vendor, and improvements arrive as software upgrades with costs attached. A laboratory that wants to classify according to its own taxonomy, on images from its own equipment, cannot get there through a supplier.

What I built

An in-house classifier trained on the laboratory's own validated images, so that the cell taxonomy, the training procedure and the evaluation criteria all stay under the laboratory's control. Owning the pipeline also makes error analysis possible — you can see which cells the model gets wrong and why, which a vendor system does not let you do.

The framing throughout was decision support, not replacement. The goal is a more reliable and more consistent first pass, so that a qualified reviewer's attention concentrates on the difficult and clinically interesting cells rather than being spent uniformly across every image.

The data

No public dataset. Every image comes from the laboratories' own routine activity, and only images that had passed human validation were retained: 39,037 images in total.

The original classification distinguishes a large number of categories, several of which correspond to successive maturation stages of the same cell line, with few examples each and boundaries that are hard to draw. Keeping them as independent classes would have deepened the imbalance and made learning harder for no clinical gain. Working with the biologists, I defined an operational taxonomy of 12 target classes, balancing biological relevance, usefulness in the laboratory's workflow, and the amount of data actually available.

Choosing the model

The objective was never the highest possible score. It was the best balance of accuracy, compute cost and inference speed — a model light and fast enough to actually run inside a laboratory's workflow.

I benchmarked eight configurations on the same test split with the same metrics: a YOLO-based classifier, two large self-supervised and vision-language backbones, a very large foundation model, plus ensemble and test-time-augmentation variants.

The compact YOLO classifier gave the best macro F1 of any single-pass model, at a fraction of the parameters of the largest architectures. Scaling the model up did not improve results on this data — the 431M-parameter foundation model finished last.

Benchmarking against a competition solution

Comparing against a public cytology data challenge's leaderboard would not have been rigorous — different dataset, different taxonomy, different evaluation protocol, and a final score that doesn't correspond to the metrics used here.

So instead of comparing numbers, I took the code of the solution that placed third in that challenge, kept its core training techniques, retrained it on the internal dataset, and evaluated it on the same test partition with the same metrics. That makes the comparison mean something: same data, same classes, same measurement.

It came out ahead by 2.18 points of accuracy and 1.76 points of macro F1. One adaptation and one training run each, so it is a directional result rather than a statistically established one — which is why the eight-architecture comparison above matters more.

Results

96.32% accuracy and 95.77% macro F1 across the twelve classes, on held-out real images.

Inference cost

Latency decides whether any of this is deployable, so all configurations were timed in the same environment.

The selected model is both the fastest single model and the best-scoring one — a rare and convenient result. The ensemble and the test-time-augmentation variant edge it out on macro F1 by 0.05 and 0.03 points respectively, and cost 3.5× and 6× the latency to do it. That is not a trade worth making for a laboratory workflow, so the single model was selected.