# SC-MAPS — Bacterial Microscopy Dataset for Deep Learning-Based Species Identification

Relevant code for this dataset: https://github.com/KapanidisLab/SC-MAPS

**Dataset creator:** Anuj Tiwari  
**Institution:** University of Oxford, United Kingdom  
**Repository:** Oxford University Research Archive (ORA)  
**Dataset title:** SC-MAPS: A Multichannel Single-Cell Microscopy Dataset for Deep Learning-Based Bacterial Species Identification and Antimicrobial Susceptibility Testing Routing  
**Archive:** `Trillian_DATASET_with_holdout.zip`

## 1. DATASET OVERVIEW

SC-MAPS is a multichannel, single-cell microscopy dataset developed to investigate whether bacterial morphology and fluorescence-staining patterns can support deep-learning-based bacterial species identification. It also supports evaluation of the assignment of predicted species to eight predefined, antimicrobial-resistance-informed groups relevant to antimicrobial susceptibility testing (AST) workflows. These AST-routing groups are a research classification framework and **do not provide an individual isolate's antibiotic susceptibility or resistance result**.

The core dataset comprises **916,435 segmented bacterial cells**, representing **13 bacterial species**, **49 species–strain/preparation combinations**, and **697 microscopy fields**. Each cell has five spatially matched imaging channels. The deposited core archive includes independent holdout data for evaluation on strains not included in model development.

## 2. DATA DESCRIPTION

### 2.1 Species represented

The main dataset contains the following 13 species (SC-MAPS abbreviations in parentheses):

1. *Acinetobacter baumannii* (ACAA)
2. *Enterobacter cloacae* complex (ENTC)
3. *Escherichia coli* (ECOL)
4. *Klebsiella aerogenes* (KLAE)
5. *Klebsiella pneumoniae* (KLPN)
6. *Pseudomonas aeruginosa* (PSAR)
7. *Salmonella enterica* (SAENA)
8. *Serratia marcescens* (SERM)
9. *Stenotrophomonas maltophilia* (PSMA)
10. *Staphylococcus aureus* (SAUR)
11. *Staphylococcus capitis* (STCP)
12. *Staphylococcus epidermidis* (SEPI)
13. *Staphylococcus lugdunensis* (SLUG)

The separately prepared dataset of additional, previously unseen organisms is **not part of this core 13-species ZIP archive** and is intended for a separate deposit.

### 2.2 Imaging channels and acquisition

Samples were imaged on the **Trillian wide-field microscope system**, built around a **ZEISS Axio Observer.Z1/7** inverted microscope with a **Plan-Apochromat 63×/1.40 NA oil-immersion Ph3 objective**, an **Axiocam 506m** camera and a 1× camera adapter. Acquisition used 1 × 1 binning, producing an effective sample-plane sampling of approximately **0.072 µm/pixel**. Images were acquired using **ZEN 2.6 (blue edition; version 2.6.76.00)**.

Five matched channels were collected:

- **Phase contrast:** transmitted-light cellular morphology.
- **Bright field (BF):** complementary transmitted-light morphology.
- **DAPI:** DNA/nucleoid-associated fluorescence.
- **WGA–Alexa Fluor 647 (WGA-AF647):** cell-envelope-associated fluorescence.
- **FM1-43:** membrane-associated fluorescence.

The fluorescence channels were acquired by wide-field epifluorescence. The manuscript documents nominal excitation/emission settings of **353/465 nm (DAPI)**, **653/668 nm (WGA-AF647)** and **473/579 nm (FM1-43)**. Acquisition exposure times differed across channels and imaging sessions; representative exposures were approximately **26–33 ms** for DAPI, bright field and phase contrast, **50–100 ms** for WGA-AF647 and **20 ms** for FM1-43. These are representative settings, not fixed values guaranteed for every image.

### 2.3 Single-cell processing

Spatially matched images were aligned by common sample and field identifiers. **Cellpose** segmentation of phase-contrast images generated cell-instance masks, which were applied consistently across the five channels. Segmented objects outside the **40–5,000-pixel** area range were excluded. The bounding box for each retained object was expanded by **8 pixels** per side, and matching regions were extracted from all channels.

Each cell was stored as a **five-channel, 5 × 96 × 96-pixel NumPy array**, in the following fixed channel order:

`PhaseContrast, BF, DAPI, WGA-AF647, FM1-43`

Crops retained native image scale, were centred and zero-padded where required, and **were not resized**. Pixels outside the focal-cell segmentation mask were excluded from the analysis crops. Crops that exceeded the 96 × 96 canvas were excluded rather than rescaled. The analysis dataset therefore contains **processed single-cell image stacks**, with cell-level metadata describing the original images and measured phenotypes; it should not be assumed to constitute an archive of all full-field raw microscope acquisitions.

### 2.4 Metadata and dataset partitions

The accompanying `cell_manifest.csv` associates image files with species labels, plate/strain assignments, source-image identifiers, segmentation-instance identifiers, morphological measurements and channel-specific intensity information. Its recorded features include cell area, axis lengths, maximum Feret diameter, aspect ratio, eccentricity, solidity, compactness and fluorescence/transmitted-light measurements.

The established dataset split is:

| Partition | Cells | Purpose |
|---|---:|---|
| Training | 625,236 | Model fitting |
| Validation | 69,473 | Model development and selection |
| Holdout (Plate 1) | 221,726 | Independent strain-level evaluation |
| **Total** | **916,435** | |

For each species, **Plate 1** was reserved for holdout testing. The remaining usable strains/preparations (Plates 2–4) formed the development dataset. Cells originating from the same microscopy field were kept within a single partition. Holdout cells must **not** be mixed into training or validation during reanalysis.

## 3. RESEARCH USE

This dataset supports research in bacterial image analysis, single-cell phenotyping, microscopy-based species identification, evaluation of deep-learning models (including single-channel and multichannel approaches), and classification into AST-relevant organism groups.

The images and labels support computational experiments; the dataset alone **does not establish phenotypic antimicrobial susceptibility, minimum inhibitory concentrations, clinical diagnostic validity or suitability for clinical decision-making**. Users should account for strain/preparation effects, sample-level dependence and the independent holdout partition when evaluating generalisation.

## 4. FILE ACCESS AND DEPOSIT

The main data are distributed in the separate compressed archive **`Trillian_DATASET_with_holdout.zip`**, approximately **6.3 GB**. For this ORA deposit, the README is provided through Symplectic Elements and the large ZIP is supplied through the **ORA Large File Upload service**, according to the repository's instructions.

The source analysis workflow identifies the following principal components within the extracted dataset:

```text
Trillian_DATASET_with_holdout/
├── cell_manifest.csv
├── splits/
│   ├── train.csv
│   ├── val.csv
│   └── holdout_plate1.csv
└── [per-cell image subdirectories]/
    └── *.npy
```

The layout above identifies documented filenames and file classes; the **exact internal subdirectory tree and file inventory must be checked against the deposited ZIP**. In particular, the location and naming of the individual `.npy` files may be nested differently.

**File formats:** `.zip` (distribution archive), `.npy` (NumPy single-cell multichannel image arrays) and `.csv` (manifest and split tables).

**Storage note:** An earlier extraction of this named ZIP on Oxford ARC reported approximately **175 GB uncompressed** and around **1.83 million archive members**. Users should provide substantial free storage before extracting; exact figures may differ if the archived version changes.

## 5. ASSOCIATED PUBLICATION, SUPPORT AND REUSE

**Associated project:** SC-MAPS — multichannel single-cell microscopy and deep-learning-based bacterial species identification and AST-routing research, University of Oxford.

**Associated publication:** SC-MAPS research manuscript (in preparation/submission at the time of this README). The publication citation and DOI should be added to the ORA record when available.

**Facilities and computing:** Imaging was supported by the MICRON Advanced Bioimaging Facility, supported by Wellcome Strategic Awards **091911/B/10/Z** and **107457/Z/15/Z**. Computational work benefited from the **University of Oxford Advanced Research Computing (ARC)** facility. Any additional research-project-specific funding should be recorded in the ORA deposit metadata or associated publication.

**Licence and reuse conditions:** The applicable licence and any access restrictions are those displayed on the final **ORA dataset record**. No specific open-data licence is asserted here before confirmation by the depositor and repository.

**Contact:** Anuj Tiwari, University of Oxford. Please use the contact details associated with the ORA deposit record.

## 6. CITATION

When using or referring to this dataset, please cite the **persistent DOI assigned by the Oxford University Research Archive**. The recommended dataset citation is:

Tiwari, A. *SC-MAPS: A Multichannel Single-Cell Microscopy Dataset for Deep Learning-Based Bacterial Species Identification and Antimicrobial Susceptibility Testing Routing*. Oxford University Research Archive. [ORA DOI to be supplied upon publication].

Users should additionally cite the associated research publication when available.