What the framework is
Four parts, each written once and reused: a data layer that turns any folder or label file into one table of images, labels and patient ids; a split builder that is deterministic and groups images by patient, so no patient sits on both sides of a split, and that respects an official test set when one exists; a per-class cap that keeps the largest classes from drowning the rest; and a training layer that runs every backbone under the same protocol and logs everything to Weights & Biases.
The backbones cover both families. Convolutional: YOLO classification models. Vision transformers: SigLIP from Google, DINOv2 from Meta and C-RADIOv4 from NVIDIA — all benchmarked on the leukocyte work. The four datasets below were run with YOLO, the strongest single model on that benchmark, and SigLIP, with the recipe left exactly as it was.
Four datasets, one recipe
Each card compares our two models with the published results on the same kind of test: an external or official test set where the dataset has one, a held-out split otherwise. Published numbers are shown only where a source states them; where a protocol is not comparable, the card says so.
One recipe, four public medical datasets, no dataset-specific tuning
Our two models against the published results on each dataset. SigLIP YOLO published
Bone marrow cytology
20 cell types · 171,366 cells · 17,137 held out
Balanced accuracy (mean recall over classes)
- SigLIP79.76%
- YOLO79.48%
- ResNeXt-50 · dataset paper, 202171.4%
- Region-attention · 2024, unseen test32%
+8.4 points over the dataset paper, trained on 30% of the images.
Matek et al., Blood 2021 · Tarquino et al., J Pathol Inform 2024
Colorectal histology
9 tissue classes · external test on 7,180 patches of new patients
Accuracy on CRC-VAL-HE-7K
- DINO · self-supervised ViT95.9%
- SigLIP95.74%
- YOLO95.32%
- VGG19 · dataset paper, 201994.3%
- DenseNet92.9%
Above the dataset paper and classic CNNs, within 0.2 points of DINO, on a third of the data.
Kather et al., PLoS Med 2019 · leaderboard compiled in Ignatov & Malivenko 2024
Retinal OCT
CNV · DME · drusen · normal · official test set of 968 scans
Accuracy on the official test set
96.49%YOLO
What we found in the official split
27,693 of the 83,516 official training images come from patients who are also in the test set. We removed them before training; results reported on the raw split include that overlap.
Counted from the patient id in every file name · Kermany et al., Cell 2018
Breast ultrasound (BUSI)
benign · malignant · normal · 780 images · 78 held out
Three-class accuracy
- DCUNet + DenseNet + LSTM · 2024(protocol not stated)99.86%
- SigLIP87.18%
- YOLO87.18%
- InceptionV3 · 202583%
- Graph transformer · 202668.8%
Above the two published results whose protocol is stated; the dataset itself ships duplicates.
ArunKumar et al., Sensors 2025 · Ali et al., Bioengineering 2026 · Meena & Joe Prathap, JBSD 2024
| Dataset | Test set | Our best | Published reference |
|---|---|---|---|
| Bone marrow cytology (20 classes) | 17,137 held-out cells | 79.76% bal. acc. | 71.4% · dataset paper |
| Colorectal histology (9 classes) | 7,180 patches, new patients | 95.74% acc. | 94.3% · dataset paper |
| Retinal OCT (4 classes) | official test, 968 scans | 96.49% acc. | test patients removed from training |
| Breast ultrasound (3 classes) | 78 held-out images | 87.18% acc. | 83.0% · 68.8% · 99.86% (protocol not stated) |
What each dataset taught
Bone marrow: the largest public cytology set, 20 cell types from 8 to 29,424 images each. Balanced accuracy — the mean recall over classes — is 8.4 points above the dataset paper's model, trained on 30% of the images; a 2024 method reports 32% on an unseen test set.
Colorectal histology: a random split of the training set scores 99.7–99.8%, because patches of the same slide land on both sides. On the official external set of new patients the same models score 95.3–95.7% — above the dataset paper and classic CNNs, within 0.2 points of a self-supervised ViT.
Retinal OCT: before training, the framework's patient grouping flagged that 27,693 of the 83,516 official training images come from patients who are also in the official test set. With them removed, the model still reaches 96.49% on the official test.
Breast ultrasound: 780 images, published accuracies from 68.8% to 99.86% depending on the protocol, and a release documented to contain duplicates. Both of our models land at 87.18%, above the two results whose protocol is stated — a reminder that on small datasets the protocol moves the number more than the model.
Code, demos and weights
Each dataset has a public write-up with figures, a live demo, and the trained weights on Hugging Face: bone marrow — repository · demo; colorectal histology — repository · demo; retinal OCT — repository · demo; breast ultrasound — repository · demo.
Limits
One training run per model and dataset, so no confidence intervals. Inputs are 224 pixels. Where a dataset has no patient ids, the split is per image. Only YOLO and SigLIP were run on these four datasets; the other backbones were benchmarked on the leukocyte data only.