DFCAN — PyTorch Implementation for Fluorescence Microscopy Super-Resolution

A clean, config-driven PyTorch re-implementation of DFCAN (Fourier Channel Attention Network) for biological image super-resolution on the BioSR dataset.
Supports 2×, 4×, and 8× upscaling on F-actin and other BioSR structures.

---

Background

DFCAN was originally proposed in:

Nat. Methods 2021 — Evaluation of deep learning strategies for nucleus segmentation in fluorescence images
Architecture reference: Wang et al., ECCV 2020 — Fourier Channel Attention Networks

This repository provides a from-scratch PyTorch port of DFCAN with a modular, config-based design. The original codebase was TensorFlow-based with hard-coded paths and deprecated APIs. This implementation is reproducible, system-independent, and built around YAML configs for easy scale/dataset switching.

---

Results

Trained on F-actin (SIM) from the BioSR dataset:

Scale	MAE	PSNR
2×	0.0465	24.74 dB

Evaluated on the BioSR validation split. Training was run on Mac (CPU) with a limited iteration budget — results are from an early checkpoint and not fully converged.

---

Repository Structure

DFCAN_PyTorch/
│
├── configs/
│   ├── dfcan_2x.yaml        # Config for 2× super-resolution
│   ├── dfcan_4x.yaml        # Config for 4× super-resolution
│   └── dfcan_8x.yaml        # Config for 8× super-resolution
│
├── src/
│   ├── models/              # DFCAN network definition
│   ├── trainers/
│   │   └── train.py         # Training loop
│   └── inference/
│       └── evaluate.py      # Evaluation & inference
│
├── data_agmt_matlab/        # MATLAB scripts for BioSR LR-HR pair generation
├── dataset/                 # Dataset root (populated after MATLAB preprocessing)
├── checkpoints/             # Saved model checkpoints
│
├── train.sh                 # Training helper script
├── eval.sh                  # Evaluation helper script
├── infer.sh                 # Inference helper script
├── requirements.txt
└── README.md

---

Installation

git clone https://github.com/Jamijunky/DFCAN_PyTorch.git
cd DFCAN_PyTorch
pip install -r requirements.txt

Requirements: Python ≥ 3.8, PyTorch ≥ 1.12, torchvision, numpy, Pillow, PyYAML, tqdm, scikit-image, matplotlib

---

Dataset Preparation

This implementation uses the BioSR dataset (F-actin structure).

1. Download BioSR from the official source.
2. Run the MATLAB preprocessing script to generate LR-HR patch pairs:

% Inside MATLAB, from the repo root:
cd data_agmt_matlab
run generate_pairs.m

This populates dataset/train/ and dataset/val/ with LR-HR .tif pairs at your chosen scale factor.

---

Training

Via shell script (Linux/Mac):

./train.sh configs/dfcan_2x.yaml
./train.sh configs/dfcan_4x.yaml
./train.sh configs/dfcan_8x.yaml

Cross-platform (Windows compatible):

python -m src.trainers.train --config configs/dfcan_4x.yaml

Checkpoints are saved to checkpoints/dfcan_4x/ (or the path specified in the config). Training resumes automatically from the latest checkpoint if one exists.

---

Evaluation

./eval.sh configs/dfcan_4x.yaml checkpoints/dfcan_4x/best.pth test

Or manually:

python -m src.inference.evaluate \
  --config configs/dfcan_4x.yaml \
  --checkpoint checkpoints/dfcan_4x/best.pth \
  --split test

Reports MAE and PSNR on the specified split.

---

Inference

./infer.sh configs/dfcan_4x.yaml checkpoints/dfcan_4x/best.pth

Output .tif files are saved to the results directory defined in the config. Scientific .tif format is used throughout to preserve image fidelity (no lossy compression).

---

Config Structure

Each YAML config controls all training and inference parameters:

scale_factor: 4
input_channels: 1         # 1 for WF, 9 for SIM
patch_size: 128
batch_size: 4
max_iters: 12000
lr: 5.0e-5
lr_decay: 0.5

data:
  train_dir: dataset/train/F-actin
  val_dir: dataset/val/F-actin

checkpoint:
  save_dir: checkpoints/dfcan_4x
  save_interval: 500

results_dir: results/dfcan_4x

---

Key Design Decisions

Why PyTorch over the original TF code?
The original DFCAN repo used TensorFlow 1.x-era patterns with deprecated keras imports, hard-coded paths, and broken scikit-image metric calls (compare_psnr, compare_ssim were removed in scikit-image ≥ 0.19). This port modernises the full stack.

Why not directly scale to 4× or 8× from a 2× checkpoint?
Naively changing scale_factor without retraining causes the model to hallucinate structures at unlearned frequency bands — the network was never constrained against those artifacts. Each scale requires either its own training run or a properly staged recursive inference pipeline (2× → 4× → 8×), which is what infer.sh implements.

Config-driven design
All hyperparameters, paths, and scale factors live in YAML configs. No hard-coded values anywhere in the source.

---

Citation

If you use this code, please cite the original DFCAN paper:

@inproceedings{wang2020fourier,
  title     = {Fourier Channel Attention Networks},
  author    = {Wang, Zhendong and others},
  booktitle = {ECCV},
  year      = {2020}
}

And the BioSR dataset:

@article{chen2021three,
  title   = {Three-dimensional residual channel attention networks denoise and sharpen fluorescence microscopy image volumes},
  author  = {Chen, Joanna and others},
  journal = {Nature Methods},
  year    = {2021}
}

---

Author

Abdullah Jami
B.Tech (IT), IIIT Sonepat
Research conducted under Dr. Hazique Aetesam, Asst. Professor, CS&E, BIT Mesra (Oct 2025 – Feb 2026)