Federated Learning for Breast Cancer Classification

Table of Contents

• Overview
• Installation
• Usage
• Dataset
• Model Architecture
• Training
• Evaluation
• Results

Overview

The project aims to classify mammograms using Deep Learning in a Federated Learning setting. By collaborating with multiple institutions, training of a central model can leverage on private data kept in the institutions, moving only the model and not the data. Specifically, the project aims to:

1. Implement a classification model for mammograms.
2. Incorporate the classification model into a Federated Learning framework.
3. Conduct experiments to assess the impact of Federated Learning on classification performance.

Installation

To clone the library and install the required libraries, use the following commands:

git clone https://github.com/e-mny/mammo_classification.git
pip install -r requirements.txt

Export the base directory to your PYTHONPATH

export PYTHONPATH="$PYTHONPATH:/path/to/base_dir"

Usage (Centralized Training)

To execute training using ResNet50 on the CBIS-DDSM dataset, use the following command:

python main.py --model resnet50 --dataset CBIS-DDSM --num_epochs 100 --data_augment --early_stopping

Option	Description	Examples
--model	Selection of model	--model efficientnet_b0 --model densenet121 --model mobiletnet_v2
--dataset	Selection of dataset	--dataset CMMD --dataset RSNA --dataset VinDr
--data_augment	Flag for data augmentation	--data_augment --no-data_augment
--early_stopping	Flag for early stopping	--early_stopping --no-early_stopping

If using SLURM, use:

sbatch cbis-resnet-slurm-job 100 true

First argument (100) is the number of epochs

Second argument (true) indicates data augmentation

Usage (Federated Learning)

To execute training of an FL model on the same machine, run the following code:

python ./flwr/server.py
python ./clients/CBIS_client.py
python ./clients/CMMD_client.py
python ./clients/RSNA_client.py
python ./clients/VinDr_client.py

Performances of each FL run can be found in the ./flwr/logs folder.

Visualization of the performance is done manually, copying the values into /flwr/graph/displayGraph.py before running the Python script.

The images of the graphs will then be saved in the /flwr/performance_metrics folder

Dataset

Refer to Datasets.MD.

For data augmentation properties, refer to data_augment.py.

Model Architecture

For our experiments, we used ResNet50 for training as we found it to have the best performance out of the state-of-the-art (SOTA) networks (Xception, MobileNet_v2, DenseNet121, EfficientNet_b0).

To test this code using other networks, refer to modelFactory.py for

• available models
• editing of classifier layer
• controlling which layers to freeze

Training

The model was trained using a batch size of 512 for 100 epochs with a learning rate of 0.001. We utilized transfer learning with a pre-trained ResNet50 model as the base and fine-tuned it on our dataset.

Training file can be found train_loader.py.

Evaluation

The model achieved the following metrics on the test set, and metrics included accuracy, precision, recall, F1-score and AUC.

Accuracy: 0.6258

Precision: 0.6503

Recall: 0.7248

F1-Score: 0.6691

AUC: 0.5951

Results

Results are saved in the Results folder, formatted as MM-DD-YYYY_HHMMSS_{dataset}_{model}.jpg.

Visualization of GradCAM on sample images can be found in Visualization folder, formatted as MM-DD-YYYY_HHMMSS.