Overview

This repository contains the code for the master thesis "Self-Supervised Representation Learning for Early Breast Cancer Detection in Mammographic Imaging" available here.

Installation

conda create --name <env> --file requirements.txt

The training code assumes Wandb is setup and logged in in the shell environment where you are running the script. If you do not want to use Wandb, comment out the Wandb call from the code.

Reproduce results

1. Downlaod RSNA and CMMD data sets
2. Preprocess the DICOM images to the same size by cropping to the ROI by using the preprocess_rsna.ipynband preprocess_cmmd.ipynb notebooks. Update the paths in the first cell of the notebooks.
3. Train SimSiam and SimClr on the RSNA dataset with and without mixup using main_parallel.py
4. Evaluate embeddings with t-SNE using evaluate_tsne.ipynb
5. Transfer weights and fine tune on the CMMD data set using evaluate_finetune_transfer.ipynb

Training

The code in main_parallel.py supports multi GPU training and will automatically use all available GPUs, and supports a number of parameters:

Parameter	Values	Description
train_path	string	Path of the train.csv file from the RSNA dataset
images_path	string	Path of the preprocessed images from the RSNA dataset
batch_size	int	Batch size used during training per GPU
simclr	store_true	Set this flag to train SimCLR, if not set, SimSiam will be trained
mixup	store_true	Set this flag to train using the Mixup variant of SimClr or SimSiam
model_type	resnet18 or resnet50	The backbone model
epochs	int	Amount of epochs to train for

Examples

SimClr with mixup:

python main_parallel.py -t data/rsna_bcd/data/train.csv -i /data/rsna_bcd_preprocessed_768_384/ -b 160 --model_type resnet18 --epochs 91 --mixup --simclr

SimSiam with mixup:

python main_parallel.py -t data/rsna_bcd/data/train.csv -i /data/rsna_bcd_preprocessed_768_384/ -b 32 --model_type resnet50 --epochs 91 --mixup