Counting 3,367 Big Data & Machine Learning Frameworks, Toolsets, and Examples...
Suggestion? Feedback? Tweet @stkim1

Last Commit
Dec. 6, 2018
Aug. 18, 2017

End-to-end Training for Whole Image Breast Cancer Screening using An All Convolutional Design

Li Shen, Ph.D. CS

Icahn School of Medicine at Mount Sinai

New York, New York, USA



This is the companion site for our paper - "End-to-end Training for Whole Image Breast Cancer Diagnosis using An All Convolutional Design" at arXiv. Access the paper here. This work was presented at the NIPS17 workshop on machine learning for health. Access the 4-page short paper here. Download the poster.

For our entry in the DREAM2016 Digital Mammography challenge, see this write-up. This work is much improved from our method used in the challenge.

Whole image model downloads

A few best whole image models are available for downloading at this Google Drive folder. YaroslavNet is the DM challenge top-performing team's method. Here is a table for individual downloads:

Database Patch Classifier Top Layers (two blocks) Single AUC Augmented AUC Link
DDSM Resnet50 [512-512-1024]x2 0.86 0.88 download
DDSM VGG16 512x1 0.83 0.86 download
DDSM VGG16 [512-512-1024]x2 0.85 0.88 download
DDSM YaroslavNet heatmap + max pooling + FC16-8 + shortcut 0.83 0.86 download
INbreast VGG16 512x1 0.92 0.94 download
INbreast VGG16 [512-512-1024]x2 0.95 0.96 download
  • Inference level augmentation is obtained by horizontal and vertical flips to generate 4 predictions.
  • The listed scores are single model AUC and prediction averaged AUC.
  • 3 Model averaging on DDSM gives AUC of 0.91
  • 2 Model averaging on INbreast gives AUC of 0.96.

Patch classifier model downloads

Several patch classifier models (i.e. patch state) are also available for downloading at this Google Drive folder. Here is a table for individual download:

Model Train Set Accuracy Link
Resnet50 S10 0.89 download
VGG16 S10 0.84 download
VGG19 S10 0.79 download
YaroslavNet (Final) S10 0.89 download
Resnet50 S30 0.91 download
VGG16 S30 0.86 download
VGG19 S30 0.89 download

With patch classifier models, you can convert them into any whole image classifier by adding convolutional, FC and heatmap layers on top and see for yourself.

A bit explanation of this repository's file structure

  • The .py files under the root directory are Python modules to be imported.
  • You shall set the PYTHONPATH variable like this: export PYTHONPATH=$PYTHONPATH:your_path_to_repos/end2end-all-conv so that the Python modules can be imported.
  • The code for patch sampling, patch classifier and whole image training are under the ddsm_train folder.
  • is used to sample patches from images and masks.
  • is used to train a patch classifier.
  • is used to train a whole image classifier, either on top of a patch classifier or from another already trained whole image classifier (i.e. finetuning).
  • There are multiple shell scripts under the ddsm_train folder to serve as examples.

Transfer learning is as easy as 1-2-3

In order to transfer a model to your own data, follow these easy steps.

Determine the rescale factor

The rescale factor is used to rescale the pixel intensities so that the max value is 255. For PNG format, the max value is 65535, so the rescale factor is 255/65535 = 0.003891. If your images are already in the 255 scale, set rescale factor to 1.

Calculate the pixel-wise mean

This is simply the mean pixel intensity of your train set images.

Image size

This is currently fixed at 1152x896 for the models in this study. However, you can change the image size when converting from a patch classifier to a whole image classifier.


Now you can finetune a model on your own data for cancer predictions! You may check out this shell script. Alternatively, copy & paste from here:

export NUM_CPU_CORES=4

python \
    --no-patch-model-state \
    --resume-from $RESUME_FROM \
    --img-size 1152 896 \
    --no-img-scale \
    --rescale-factor 0.003891 \
    --featurewise-center \
    --featurewise-mean 44.33 \
    --no-equalize-hist \
    --batch-size 4 \
    --train-bs-multiplier 0.5 \
    --augmentation \
    --class-list neg pos \
    --nb-epoch 0 \
    --all-layer-epochs 50 \
    --load-val-ram \
    --load-train-ram \
    --optimizer adam \
    --weight-decay 0.001 \
    --hidden-dropout 0.0 \
    --weight-decay2 0.01 \
    --hidden-dropout2 0.0 \
    --init-learningrate 0.0001 \
    --all-layer-multiplier 0.01 \
    --es-patience 10 \
    --auto-batch-balance \
    --best-model $BEST_MODEL \
    --final-model $FINAL_MODEL \

Some explanations of the arguments:

  • The batch size for training is the product of --batch-size and --train-bs-multiplier. Because training uses roughtly twice (both forward and back props) the GPU memory of testing, --train-bs-multiplier is set to 0.5 here.
  • For model finetuning, only the second stage of the two-stage training is used here. So --nb-epoch is set to 0.
  • --load-val-ram and --load-train-ram will load the image data from the validation and train sets into memory. You may want to turn off these options if you don't have sufficient memory. When turned off, out-of-core training will be used.
  • --weight-decay and --hidden-dropout are for stage 1. --weight-decay2 and --hidden-dropout2 are for stage 2.
  • The learning rate for stage 1 is --init-learningrate. The learning rate for stage 2 is the product of --init-learningrate and --all-layer-multiplier.

Computational environment

The research in this study is carried out on a Linux workstation with 8 CPU cores and a single NVIDIA Quadro M4000 GPU with 8GB memory. The deep learning framework is Keras 2 with Tensorflow as the backend.