SatellaDet-Blood

SatellaDet-Blood is a public benchmark of the custom SatellaDet object-detection architecture on the TXL-PBC peripheral blood-cell dataset.

SatellaDet was originally developed for dense microbiological object detection.

This experiment tests whether the same detector architecture can generalise to a substantially different microscopy domain without redesigning the architecture specifically for blood-cell detection.

The model was trained from scratch at 640 × 640 resolution.

It detects three classes:

WBC
RBC
Platelets

Held-out test results

The checkpoint was selected using the validation split only.

After training and checkpoint selection were complete, best_score.pt was evaluated against the untouched official TXL-PBC test split.

Metric	Held-out test
Precision	0.9239
Recall	0.9065
mAP50	0.9341
mAP50-95	0.7346
Count MAE	1.2540
Count bias	-0.6032

Per-class held-out performance

Class	GT	Pred	Precision	Recall	F1	mAP50	mAP50-95
WBC	133	111	1.0000	0.8346	0.9098	0.9172	0.7163
RBC	1,699	1,636	0.9615	0.9258	0.9433	0.9523	0.7858
Platelets	49	58	0.8103	0.9592	0.8785	0.9330	0.7018

Selected validation checkpoint

The SatellaScore-selected checkpoint was obtained at epoch 64.

Metric	Validation
Precision	0.941
Recall	0.904
mAP50	0.9378
mAP50-95	0.7353
SatellaScore	0.8602

Validation and untouched test performance were closely aligned:

Metric       Validation     Test
mAP50          0.9378      0.9341
mAP50-95       0.7353      0.7346

The official test split was not used for checkpoint selection.

SatellaScore

Checkpoint selection used equal class weighting.

This prevents the substantially more numerous RBC annotations from dominating checkpoint selection.

WBC        0.333333
RBC        0.333333
Platelets  0.333334

SatellaScore composition:

mAP50-95   35%
mAP50      25%
F1         30%
Count      10%

Included files

Models

model/SatellaDet_Blood_TXL_PBC_640.onnx

Deployable ONNX export of the selected SatellaDet-Blood model.

model/best_score.pt

PyTorch checkpoint selected using SatellaScore.

Benchmark artifacts

The results/ directory contains:

detection metric graphs
training-loss graph
count-error graph
held-out test metrics
held-out test summary
sanitised training summary

PUBLIC_MANIFEST.json contains SHA-256 hashes and file sizes for the public release.

Dataset

This benchmark uses the public TXL-PBC peripheral blood-cell dataset.

TXL-PBC contains:

1,260 microscopy images
18,143 annotated blood cells
official training, validation, and test splits

Reference:

Gan, L., Li, X. & Wang, X.

TXL-PBC: a peripheral blood cell dataset with comprehensive annotations.

Scientific Data 12, 1694 (2025).

DOI:

10.1038/s41597-025-05980-z

Dataset repository:

https://github.com/lugan113/TXL-PBC_Dataset

SatellaDet

SatellaDet is a custom object-detection architecture developed by Ephraim Asad.

Source repository:

https://github.com/EphraimAsad/SatellaDet

SatellaDet-Blood was trained from scratch rather than fine-tuned from a pretrained blood-cell detector.

The purpose of this experiment was to test whether the architecture could generalise beyond the microbiological colony-detection tasks for which it was originally developed.

Privacy

This public release contains only:

model weights
aggregate benchmark metrics
generated training graphs
public model documentation

It intentionally excludes:

raw microscopy images
raw TXL-PBC dataset files
private laboratory data
training console logs
environment dumps
notebooks
local configuration files
API credentials
user-specific filesystem information

Intended use

SatellaDet-Blood is intended for:

object-detection research
computer-vision experimentation
benchmarking
architecture evaluation
educational use

It is not a clinical diagnostic system.

It should not be used to make medical diagnoses or patient-care decisions.

Benchmark note

The headline metrics in this repository correspond to the untouched TXL-PBC test split.

Because training settings and evaluation implementations can differ between detectors, comparisons with results published by other architectures should be treated as contextual unless evaluated under a matched experimental protocol.

Downloads last month: 18