DNA Image Classification Pipeline

A deep learning pipeline for classifying DNA morphological structures from microscopy images. The system uses a ResNet-50 backbone with a two-layer MLP head, fine-tuned with layer-wise learning rates and balanced class sampling.

Features

• Object cropping: Automatic segmentation of individual DNA objects from full-field microscopy images using binary thresholding and connected-component analysis
• Aspect-ratio-preserving resize: Custom ResizeWithPad transform that pads with ImageNet mean colour
• Balanced training: Oversampling of minority classes via BalancedClassSampler
• Layer-wise learning rates: Differential learning rates for early, middle, and late backbone layers
• OOD detection: Confidence-based out-of-distribution filtering at inference time
• Interactive UI: Gradio web interface for the full labelling → training → inference loop

Installation

pip install -r requirements.txt

Usage

Training (CLI)

Prepare your data in ImageFolder format:

data/train/
├── ClassA/
│   ├── img001.png
│   └── ...
├── ClassB/
│   └── ...

Then run:

python train.py --data_root ./data/train --epochs 13 --gpu 0

Inference (CLI)

python inference.py \
    --model_path ./checkpoints/model.pth \
    --classes_path ./checkpoints/classes.txt \
    --data_root ./data/test \
    --confidence_threshold 96 \
    --copy_images

Gradio Web Interface

python app.py --port 7860

Model Architecture

Component	Details
Backbone	ResNet-50 (ImageNet V2 pretrained)
Head	Linear(2048, 4096) → ReLU → Dropout(0.7) → Linear(4096, C)
Optimizer	AdamW with layer-wise LR
Loss	Cross-entropy
Input size	224 × 224 (aspect-ratio preserving pad)

Layer-wise Learning Rates

Layer group	Learning rate
conv1, bn1, layer1 (early)	7.3 × 10⁻⁷
layer2 (middle)	4.1 × 10⁻⁶
layer3, layer4 (late)	5.1 × 10⁻⁶
MLP head	2.5 × 10⁻⁵

Project Structure

├── train.py            # CLI training script
├── inference.py        # CLI inference script
├── app.py              # Gradio web demo
├── requirements.txt
└── src/
    ├── model.py        # Classifier and MLPHead
    ├── data.py         # Transforms, datasets, samplers
    └── preprocess.py   # Microscopy image cropping

Citation

@article{TODO,
  title={TODO},
  author={TODO},
  year={2025}
}

License

This project is licensed under the MIT License.