SCI-AIDE : High-fidelity Few-shot Histopathology Image Synthesis for Rare Cancer Diagnosis

Pretrained Models

In this work, we created synthetic tissue microscopy images using few-shot learning and developed a digital pathology pipeline called SCI-AIDE to improve diagnostic accuracy. Since rare cancers encompass a very large group of tumours, we used childhood cancer histopathology images to develop and test our system. Our computational experiments demonstrate that the synthetic images significantly enhances performance of various AI classifiers.

Example Results

Real and Synthetic Images

Dataset

In this study, we conducted experiments using histopathological whole slide images(WSIs) of five rare childhood cancer types and their sub-types, namely ependymoma (anaplastic, myxopapillary, subependymoma and no-subtype), medulloblastoma (anaplastic, desmoplastic and no-subtype), Wilms tumour, also known as nephroblastoma (epithelial, blastomatous, stromal, Wilms epithelial-stromal, epithelial-blastomatous and blastomatous-stromal), pilocytic astrocytoma and Ewing sarcoma.

Tumour histopathology WSIs are collected at Ege University, Turkey and Aperio AT2 scanner digitised the WSIs at 20× magnification. WSIs will be available publicly soon

Prerequisites

• Linux (Tested on Red Hat Enterprise Linux 8.5)
• NVIDIA GPU (Tested on Nvidia GeForce RTX 3090 Ti x 4 on local workstations, and Nvidia A100 GPUs on TRUBA
• Python (3.9.7), matplotlib (3.4.3), numpy (1.21.2), opencv (4.5.3), openslide-python (1.1.1), openslides (3.4.1), pandas (1.3.3), pillow (8.3.2), PyTorch (1.9.0), scikit-learn (1.0), scipy (1.7.1), tensorboardx (2.4), torchvision (0.10.1).

Getting started

• Clone this repo:

git clone https://github.com/ekurtulus/SCI-AIDE.git
cd SCI-AIDE

• Install PyTorch 3.9 and other dependencies (e.g., PyTorch).

• For pip users, please type the command pip install -r requirements.txt.

• For Conda users, you can create a new Conda environment using conda env create -f environment.yml.

Synthetic Images Generation

• Clone FastGAN repo:

git clone https://github.com/odegeasslbc/FastGAN-pytorch.git
cd FastGAN-pytorch

• Train the FastGAN model:

python classifer.py --path $REAL_IMAGE_DIR --iter 100000 --batch_size 16

• Inference the FastGAN model:

python eval.py --ckpt $CKPT_PATH --n_sample $NUMBERS_OF_SAMPLE

• Train the SCI-AIDE model:

python train.py --datapath $DATAPATH_PATH --model $MODEL --savepath $SAVING_PATH --task $TRAINING_TASK

The list of other arguments is as follows:

• --lr : Learning rate (default: 5e-5)

• --opt : Optimizers ( "Adam", "SGD", "RMSprop", "AdamW" , default= "SGD")

• --batch-size : Batch size (default: 32)

• --halftensor : Mixed presicion acivaiton

• --epochs : Numbers of epochs

• --scheduler : Learning scheduler ( "cosine", "multiplicative" , default="cosine")

• --augmentation : Augmentation selection ( "randaugment", "autoaugment", "augmix", "none", default= "randaugment" )

• --memory : Data reading selection ( "none", "cached", default= "none" )

• Evaluation the SCI-AIDE model:

python wsi_attention.py --datapath $DATAPATH_PATH --model $MODEL --model_weights $MODEL_WEIGHT --output $OUTPUT_PATH --name $NAME --num_classes $NUM_CLASSES

The list of other arguments is as follows:

• --attention_level : ("pixel", "patch", default="patch)

• --cam : CAM selection ( "GradCAM", "ScoreCAM", "GradCAMPlusPlus", "AblationCAM", "XGradCAM", "EigenCAM", "FullGrad", default="EigenCAM" )

• Diagnosis WSI with the SCI-AIDE model:

python wsi_diagnosis.py --task $DIAGNOSIS_TASK --datapath $WSI_PATH --output $OUTPUT_PATH --config $CONFIG_FILE_PATH --name $NAME

The list of other arguments is as follows:

• --overlap : Patches overlaping raito (default :0 )
• --patch_size : WSI oatching size (default : 1024 )
• --heatmap : Heatmap inference activation
• --white_threshold : White pathch elimiantion ration (default :0.3)

Apply a pre-trained SCI-AIDE model and evaluate

For reproducability, you can download the pretrained models for each algorithm here.

Issues

• Please report all issues on the public forum.

License

© This code is made available under the GPLv3 License and is available for non-commercial academic purposes.

Reference

If you find our work useful in your research or if you use parts of this code please consider citing our paper:

Acknowledgments

Our code is developed based on pytorch-image-models. We also thank pytorch-fid for FID computation, and FastGAN-pytorch for the PyTorch implementation of FastGAN used in our single-image translation setting.