README.md

metadata

license: mit
tags:
  - pytorch
  - computer-vision
  - medical-imaging
  - multi-task-learning
  - classification
  - regression
  - single-cell
  - microscopy
  - white-blood-cell
datasets:
  - BSCCM
  - BCCD
language:
  - en
pipeline_tag: image-classification

Single-Cell Phenotyping — Hybrid CNN-ViT Multi-Task Model

Paper: Towards Label-Free Single-Cell Phenotyping Using Multi-Task Learning
Authors: Saqib Nazir and Ardhendu Behera (Edge Hill University, UK)
Conference: ICPR 2026
Code: github.com/saqibnaziir/Single-Cell-Phenotyping

---

Model Description

A unified deep learning framework that jointly performs White Blood Cell (WBC) classification and continuous protein-expression regression from label-free Differential Phase Contrast (DPC) microscopy images — without fluorescent staining.

Architecture

Input: (B, 4, 28, 28)  ← 4-channel DPC (Left, Right, Top, Bottom)
    │
    └─ Shared ECA Channel Attention
           ├─ CNN Branch: Stem → Inception×2 + Residual → (B, 196, 192)
           └─ ViT Branch: Patch(4×4) → Transformer×2 → (B, 50, 128)
                  │
                  └─ Cross-Modal Fusion (256-dim, learnable weights)
                         ├─ Task-Specific Refinement
                         └─ Task Gating (bidirectional cross-task exchange)
                                ├─ Classification Head → (B, 3)
                                └─ Regression Head    → (B, 4)

• Parameters: ~12M
• FLOPs: ~0.8 GFLOPs per 28×28 image (real-time capable)
• Input: 4-channel DPC images, 28×28 pixels

Tasks

Task	Output	Classes / Markers
WBC Classification	3 classes	Lymphocyte, Granulocyte, Monocyte
Protein Regression	4 markers	CD45, CD3, CD19, CD14

---

Files in This Repository

File	Description
bsccm_best_model.pth	Best checkpoint on BSCCM dataset (91.3% accuracy)
bccd_best_model.pth	Best checkpoint on BCCD benchmark (93.77% accuracy)

---

Performance

BSCCM Dataset

Metric	Value
WBC Classification Accuracy	91.3%
Macro F1-Score	0.92
Pearson r (CD16 regression)	0.73
RMSE	0.68

BCCD Benchmark (Classification Only)

Class	Precision	Recall	F1
Lymphocyte	1.000	1.000	1.000
Granulocyte	0.889	1.000	0.941
Monocyte	1.000	0.750	0.857
Macro Avg.	0.963	0.917	0.933

Overall BCCD accuracy: 93.77%

---

Usage

Installation

git clone https://github.com/saqibnaziir/Single-Cell-Phenotyping.git
cd Single-Cell-Phenotyping
pip install -r requirements.txt

Load Model

import torch
from huggingface_hub import hf_hub_download
from model import create_model

# Download checkpoint
ckpt_path = hf_hub_download(
    repo_id="saqialii/single-cell-phenotyping",
    filename="bsccm_best_model.pth"
)

# Create model and load weights
model = create_model(num_classes=3, num_proteins=4, img_size=28, in_channels=4)
checkpoint = torch.load(ckpt_path, map_location='cpu', weights_only=False)
model.load_state_dict(checkpoint['model_state_dict'])
model.eval()

print(f"Loaded model — best val accuracy: {checkpoint['best_val_acc']:.2f}%")

Inference

import torch
import torch.nn.functional as F

# Input: (B, 4, 28, 28) — 4-channel DPC image, normalised to [-1, 1]
image = torch.randn(1, 4, 28, 28)  # replace with real image

with torch.no_grad():
    cls_logits, prot_preds = model(image)

# Cell type classification
probs = F.softmax(cls_logits, dim=1)
class_names = ['Lymphocyte', 'Granulocyte', 'Monocyte']
predicted_class = class_names[probs.argmax().item()]
confidence = probs.max().item()

print(f"Predicted: {predicted_class} ({confidence:.1%})")

# Protein expression (Z-scored)
protein_names = ['CD45', 'CD3', 'CD19', 'CD14']
for name, val in zip(protein_names, prot_preds[0].tolist()):
    print(f"  {name}: {val:.3f}")

Data Preparation

# Download BSCCM dataset
pip install bsccm
python -c "from bsccm import download_dataset; download_dataset('./data', mnist=True)"

# Train from scratch
python train.py --data_path ./data/BSCCMNIST --save_dir checkpoints/run1

# Evaluate
python evaluate.py \
    --model_path checkpoints/run1/best_model.pth \
    --data_path ./data/BSCCMNIST \
    --output_dir evaluation_results

---

Dataset

BSCCM (Berkeley Single Cell Computational Microscopy):

• 7,889 single-cell DPC images at 28×28 pixels
• 3 WBC classes: Lymphocyte (456), Granulocyte (736), Monocyte (226) — test split
• 4 protein markers measured by fluorescence: CD45, CD3, CD19, CD14
• Source: Waller-Lab/BSCCM

BCCD (Blood Cell Images):

• ~12,000 RGB images at 128×128 pixels (4-class → mapped to 3-class)
• Source: Kaggle: Blood Cell Images

---

Citation

@inproceedings{nazir2026label,
  title     = {Towards Label-Free Single-Cell Phenotyping Using Multi-Task Learning},
  author    = {Nazir, Saqib and Behera, Ardhendu},
  booktitle = {Proceedings of the International Conference on Pattern Recognition (ICPR)},
  year      = {2026},
  note      = {arXiv:2605.14717}
}

@article{pinkard2024berkeley,
  title   = {Berkeley Single Cell Computational Microscopy Dataset},
  author  = {Pinkard, Henry and others},
  journal = {arXiv preprint arXiv:2402.06191},
  year    = {2024}
}

---

License

MIT License — Saqib Nazir, Ardhendu Behera, Edge Hill University, 2026