Breast Cancer Histopathology Image Classification and Localization using Multiple Instance Learning

Official implementation of "Breast Cancer Histopathology Image Classification and Localization using Multiple Instance Learning", published at WIECON-ECE 2019.

Introduction

Deep learning in histopathology has attracted growing interest over the past decade thanks to its improvements on classification and localization tasks. Breast cancer is a prominent cause of death in women, and Computer-Aided Pathology is essential to analyze microscopic histopathology images for diagnosis given the increasing number of patients.

Convolutional neural networks give strong results in classifying cancer vs. non-cancer tissue images, but lack interpretability. In this work we frame image classification as a weakly-supervised multiple instance learning (MIL) problem and use attention over instances (patches) to localize tumour and normal regions — providing a better interpretation of the classification results. Attention-based multiple instance learning (A-MIL) is applied on the BreakHis and BACH datasets, and produces better localization without compromising classification accuracy.

Grad-CAM and its limits on histopathology

Grad-CAM and Guided Grad-CAM are among the most popular methods for visualizing which regions a CNN uses for its prediction. Using the code in grad_cam/, here is Guided Grad-CAM applied to a BreakHis patch:

Input patch

Guided Grad-CAM

On natural images Grad-CAM localizes objects cleanly, but on H&E histopathology the attribution is diffuse and noisy — it does not isolate the diagnostically-relevant glands. This motivates an attention-based approach.

Attention-based Multiple Instance Learning (A-MIL)

We crop each image into small patches to form a bag, extract per-patch (instance) features with a CNN, compute an attention weight for each patch, and aggregate the patches into a bag-level feature for classification (Benign / Malignant / Invasive):

Multiplying each patch by its attention weight and stitching the patches back together reconstructs a localization map. A-MIL highlights the affected glands and ignores the background, while Grad-CAM on the same patches stays noisy:

The same method applied to the BACH dataset (bright = attended informative regions, dark = ignored background):

Repository structure

Path	Description
my_network/	Custom CNN architecture — net.py
ResNet/	Transfer learning with a pretrained ResNet — resnet_pre.py
VGG/	Transfer learning with a pretrained VGG — vgg_pre.py
AMIL_codes/	Attention-based MIL — model (amil_model.py), patch/bag loader (patch_data.py), train + test (train_n_test.py)
grad_cam/	Grad-CAM / Guided Grad-CAM visualization toolkit
AMIL_Data/	Expected data layout for A-MIL (readme_data_format.txt)

How to run the code

Each model folder trains on the chosen magnification (zoom) level and saves the model, terminal logs, TensorBoard logs and a run summary into the corresponding zoom-level folder.

• Custom network: cd my_network && python net.py
• ResNet (pretrained): cd ResNet && python resnet_pre.py
• VGG (pretrained): cd VGG && python vgg_pre.py
• Attention MIL: cd AMIL_codes && python train_n_test.py — trains A-MIL, tests it, and saves the patch-stitched visualizations.

Each folder also contains run_for_all_zoom.sh, which loops the corresponding script over all magnifications (40X, 100X, 200X, 400X):

bash run_for_all_zoom.sh

Grad-CAM visualization: put a test image in grad_cam/inputs/, set its filename in grad_cam/src/misc_functions.py (around line 253), then run python grad_cam/src/guided_gradcam.py. Results are written to grad_cam/results/.

Dataset

BreakHis (used by the ResNet / VGG / custom-network models):

• Official source: Breast Cancer Histopathological Database (BreakHis), UFPR
• Also on Kaggle: BreakHis dataset

A-MIL data layout: for attention-based MIL, re-arrange the dataset into the structure described in AMIL_Data/readme_data_format.txt:

data_breakhis/
  40X/  100X/  200X/  400X/
    train/
      0/   (benign images)
      1/   (malignant images)
    test/
      0/   1/

Requirements

Note: this code was developed in 2019 on an older stack — PyTorch 0.4.1, torchvision 0.2.1, TensorFlow 1.13.1 (see requirement.txt). These pins no longer install on modern Python; to reproduce the original results, use a matching legacy environment (e.g. Python 3.6). Porting to a current PyTorch is straightforward but has not been re-validated here.

Links

• Article on Medium
• arXiv:2003.00823 · IEEE Xplore · Harvard ADS

Citation

@INPROCEEDINGS{9019916,
  author={Patil, Abhijeet and Tamboli, Dipesh and Meena, Sunil and Anand, Deepak and Sethi, Amit},
  booktitle={2019 IEEE International WIE Conference on Electrical and Computer Engineering (WIECON-ECE)},
  title={Breast Cancer Histopathology Image Classification and Localization using Multiple Instance Learning},
  year={2019},
  pages={1-4},
}

Plain text: A. Patil, D. Tamboli, S. Meena, D. Anand and A. Sethi, "Breast Cancer Histopathology Image Classification and Localization using Multiple Instance Learning," 2019 IEEE International WIE Conference on Electrical and Computer Engineering (WIECON-ECE), Bangalore, India, 2019, pp. 1-4.

License

Released under the MIT License.