polejowska / LCBSI

Leukocytes (WBCs) subtypes classification from blood smear images using Vision Transformers from Hugging Face and DenseNet artificial neural network from MONAI.

LCBSI

Leukocytes classification from blood smear images.

Screen.Recording.2023-01-07.at.20.06.20.mp4

Set up the environment

Create the environment: conda env create -f environment.yml

Update the environment: conda env update -f environment.yml

Run the app (in the root project directory): python index.py

Main models notebooks

https://github.com/AgataPolejowska/LCBSI/tree/main/notebooks/main_notebooks

Dataset

The dataset used is created by combining the following datasets:

• RAABINC WBC

Kouzehkanan, Zahra Mousavi, et al. "A large dataset of white blood cells containing cell locations and types, along with segmented nuclei and cytoplasm." Scientific reports 12.1 (2022): 1-14.

• PBC

Acevedo, Andrea; Merino, Anna; Alférez, Santiago; Molina, Ángel; Boldú, Laura; Rodellar, José (2020), “A dataset for microscopic peripheral blood cell images for development of automatic recognition systems”, Mendeley Data, V1, doi: 10.17632/snkd93bnjr.1

Data is split to 70% training data, 15% validation data and 15% test data.

5000 images are divided into:

• train data: 3500 images - 700 images per each class (350 for each dataset except basophil class from PBC - 550 images and from RAABIN-WBC - 150 images)
• validation data: 750 images - 150 images per each class (75 for each dataset except basophil class from PBC - 45 images and from RAABIN-WBC - 30 images)
• test data: 750 images - 150 images per each class (75 for each dataset except basophil class from PBC - 45 images and from RAABIN-WBC - 30 images)

Hugging Face Hub Dataset: https://huggingface.co/datasets/polejowska/lcbsi-wbc-ap

Models

DenseNet121

Pretrained model can be downloaded from: https://github.com/Project-MONAI/model-zoo/releases/tag/hosting_storage_v1

Dataset description that the pretrained model used is available here: https://github.com/Project-MONAI/model-zoo/tree/dev/models/pathology_nuclei_classification

Model zoo: https://github.com/Project-MONAI/tutorials/tree/main/model_zoo/transfer_learning_with_bundle

Additional information about DenseNet121: https://docs.monai.io/en/latest/networks.html#densenet121

Other models available in MONAI: https://github.com/Project-MONAI/model-zoo/tree/dev/models

Experiments

W&B runs

1. DenseNet121 MONAI + AI Lightning sweeps https://wandb.ai/polejowska/lcbsi-densenet-monai-ap

2. ViTs architectures https://wandb.ai/polejowska/vit-classification-lcbsi

3. ViTs hyperparameters sweeps https://wandb.ai/polejowska/lcbsi-vits-sweeps

Finally developed model: https://huggingface.co/polejowska/swin-tiny-patch4-window7-224-lcbsi-wbc-new

W&B reports

https://wandb.ai/polejowska/vit-classification-lcbsi/reports/Leukocytes-classification-from-blood-smear-images--VmlldzozMTU1NjI0

Additional application

You can experiment with the trained vision transformer in the Hugging Face space: https://huggingface.co/spaces/polejowska/LCBSI