A Divide and Conquor Approach to Maximise Deep Learning Mammography Classification Accuracies - Published in PLOS ONE
![Jupyter Notebook](https://camo.githubusercontent.com/c044ae9d0419850e7f2385c22ea5de56e101e6a616789bd35d2d8fa137a63642/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f6a7570797465722d2532334641304630302e7376673f7374796c653d666f722d7468652d6261646765266c6f676f3d6a757079746572266c6f676f436f6c6f723d7768697465)
Publication repository of the "A Divide and Conquor Approach to Maximise Deep Learning Mammography Classification Accuracies" peer-reviewed paper published in PLOS ONE. You can read the paper here: https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0280841
Breast cancer claims 11,400 lives on average every year in the UK, making it one of the deadliest diseases. Mammography is the gold standard for detecting early signs of breast cancer, which can help cure the disease during its early stages. However, incorrect mammography diagnoses are common and may harm patients through unnecessary treatments and operations (or a lack of treatments). Therefore, systems that can learn to detect breast cancer on their own could help reduce the number of incorrect interpretations and missed cases. Various deep learning techniques, which can be used to implement a system that learns how to detect instances of breast cancer in mammograms, are explored throughout this paper.
Convolution Neural Networks (CNNs) are used as part of a pipeline based on deep learning techniques. A divide and conquer approach is followed to analyse the effects on performance and efficiency when utilising diverse deep learning techniques such as varying network architectures (VGG19, ResNet50, InceptionV3, DenseNet121, MobileNetV2), class weights, input sizes, image ratios, pre-processing techniques, transfer learning, dropout rates, and types of mammogram projections.
Multiple techniques are found to provide accuracy gains relative to a general baseline (VGG19 model using uncropped 512x512 pixels input images with a dropout rate of 0.2 and a learning rate of 1ร10^โ3) on the Curated Breast Imaging Subset of DDSM (CBIS-DDSM) dataset. These techniques involve transfer learning pre-trained ImagetNet weights to a MobileNetV2 architecture, with pre-trained weights from a binarised version of the mini Mammography Image Analysis Society (mini-MIAS) dataset applied to the fully connected layers of the model, coupled with using weights to alleviate class imbalance, and splitting CBIS-DDSM samples between images of masses and calcifications. Using these techniques, a 5.28% gain in accuracy over the baseline model was accomplished. Other deep learning techniques from the divide and conquer approach, such as larger image sizes, do not yield increased accuracies without the use of image pre-processing techniques such as Gaussian filtering, histogram equalisation and input cropping.
@software{Jaamour_Adamouization_Breast-Cancer-Detection-Mammogram-Deep-Learning-Publication_PLOS_ONE_2023,
author = {Jaamour, Adam and Myles, Craig},
license = {BSD-2-Clause},
month = may,
title = {{Adamouization/Breast-Cancer-Detection-Mammogram-Deep-Learning-Publication: PLOS ONE Submission}},
url = {https://github.com/Adamouization/Breast-Cancer-Detection-Mammogram-Deep-Learning-Publication},
version = {1.2},
year = {2023}
}
@article{10.1371/journal.pone.0280841,
doi = {10.1371/journal.pone.0280841},
author = {Jaamour, Adam AND Myles, Craig AND Patel, Ashay AND Chen, Shuen-Jen AND McMillan, Lewis AND Harris-Birtill, David},
journal = {PLOS ONE},
publisher = {Public Library of Science},
title = {A divide and conquer approach to maximise deep learning mammography classification accuracies},
year = {2023},
month = {05},
volume = {18},
url = {https://doi.org/10.1371/journal.pone.0280841},
pages = {1-24},
number = {5},
}
Clone the repository:
git clone https://github.com/Adamouization/Breast-Cancer-Detection-Mammogram-Deep-Learning-Publication
cd Breast-Cancer-Detection-Mammogram-Deep-Learning-Publication
Create a virtual conda environment:
conda create -n mammography python=3.6.13
conda activate mammography
Install requirements:
pip install -r requirements.txt
Create output
and save_models
directories to store the results:
mkdir output
mkdir saved_models
cd
into the src
directory and run the code:
cd ./src
main.py [-h] -d DATASET [-mt MAMMOGRAMTYPE] -m MODEL [-r RUNMODE] [-lr LEARNING_RATE] [-b BATCHSIZE] [-e1 MAX_EPOCH_FROZEN] [-e2 MAX_EPOCH_UNFROZEN] [-roi] [-v] [-n NAME]
where:
-h
is a flag for help on how to run the code.DATASET
is the dataset to use. Must be eithermini-MIAS
,mini-MIAS-binary
orCBIS-DDMS
. Defaults toCBIS-DDMS
.MAMMOGRAMTYPE
is the type of mammograms to use. Can be eithercalc
,mass
orall
. Defaults toall
.MODEL
is the model to use. Must be eitherVGG-common
,VGG
,ResNet
,Inception
,DenseNet
,MobileNet
orCNN
.RUNMODE
is the mode to run in (train
ortest
). Default value istrain
.LEARNING_RATE
is the optimiser's initial learning rate when training the model during the first training phase (frozen layers). Defaults to0.001
. Must be a positive float.BATCHSIZE
is the batch size to use when training the model. Defaults to2
. Must be a positive integer.MAX_EPOCH_FROZEN
is the maximum number of epochs in the first training phrase (with frozen layers). Defaults to100
.MAX_EPOCH_UNFROZEN
is the maximum number of epochs in the second training phrase (with unfrozen layers). Defaults to50
.-roi
is a flag to use versions of the images cropped around the ROI. Only usable with mini-MIAS dataset. Defaults toFalse
.-v
is a flag controlling verbose mode, which prints additional statements for debugging purposes.NAME
is name of the experiment being tested (used for saving plots and model weights). Defaults to an empty string.
python main.py -d CBIS-DDSM -mt all -m MobileNet -r train -lr 0.0001
python main.py -d CBIS-DDSM -mt all -m MobileNet -r test -lr 0.0001
These datasets are very large (exceeding 160GB) and more complex than the mini-MIAS dataset to use. They were downloaded by the University of St Andrews School of Computer Science computing officers onto \textit{BigTMP}, a 15TB filesystem that is mounted on the Centos 7 computer lab clients with NVIDIA GPUsusually used for storing large working data sets. Therefore, the download process of these datasets will not be covered in these instructions.
The generated CSV files to use these datasets can be found in the /data/CBIS-DDSM
directory, but the mammograms will have to be downloaded separately. The DDSM dataset can be downloaded here, while the CBIS-DDSM dataset can be downloaded here.
-
This example will use the mini-MIAS dataset. After cloning the project, travel to the
data/mini-MIAS
directory (there should be 3 files in it). -
Create
images_original
andimages_processed
directories in this directory:
cd data/mini-MIAS/
mkdir images_original
mkdir images_processed
- Move to the
images_original
directory and download the raw un-processed images:
cd images_original
wget http://peipa.essex.ac.uk/pix/mias/all-mias.tar.gz
- Unzip the dataset then delete all non-image files:
tar xvzf all-mias.tar.gz
rm -rf *.txt
rm -rf README
- Move back up one level and move to the
images_processed
directory. Create 3 new directories there (benign_cases
,malignant_cases
andnormal_cases
):
cd ../images_processed
mkdir benign_cases
mkdir malignant_cases
mkdir normal_cases
- Now run the python script for processing the dataset and render it usable with Tensorflow and Keras:
python3 ../../../src/dataset_processing_scripts/mini-MIAS-initial-pre-processing.py
- see BSD 2-Clause License file.
- Adam Jaamour (adam[at]jaamour[dot]com)
- Craig Myles
- Ashay Patel
- Shuen-Jen Chen
- Lewis McMillan
- David Harris-Birtill