Introduction

This repo is an unofficial implementation, with a single GPU restriction, of <An annotation-free whole-slide training approach to pathological classification of lung cancer types by deep learning>

Visit https://github.com/aetherAI/whole-slide-cnn to see the original implementation.

Requirements

matplotlib==3.6.2
numpy==1.23.5
pandas==1.4.2
Pillow==9.2.0
python==3.9.0
scikit-learn==1.0.2
timm==0.6.12
torch==1.13.0+cu116
torchvision==0.14.0
tqdm==4.64.1
yacs==0.1.8

Usage

python main.py your/config/path.py

For example,

python main.py configs/leaf_224_train.py

"configs/leaf_224_train.py" is for Plant Pathology 2021 - FGVC8 Dataset

Implementation Scheme

Experiments

Dataset : Plant Pathology 2021 - FGVC8
Epoch : 1 (with 15833 imgs for train and 2799 imgs for valid)
The time taken : 20h (with GeForce RTX 2070 SUPER)
Learning curves (train) :

validation accuracy : 0.836
validation mean f1score : 0.901

private score : 0.79438 (late submission on Plant Pathology 2021 - FGVC8 kaggle competition)

Considerations

Batch normalization shouldn't be used because the extractor is updated via gradient accumulation. As a result, I opted for NFNet over ResNet for the extractor.

But the following could be considered:

• Batch normalization with low momentum
• Other normalizations e.g. layer normalization