Main paper to be cited
@article{graziani2018regression,
title={Regression Concept Vectors for Bidirectional Explanations in Histopathology},
author={Graziani, Mara and Andrearczyk, Vincent and Muller, Henning},
journal={Understanding and Interpreting Machine Learning in Medical Image Computing Applications: First International Workshops},
year={2018}
}
This repository contains the code for implementing Regression Concept Vectors and Bidirectional Relevance scores, which were used to obtain the results presented in the paper.
This code runs in Python >= 2.7. Keras >= 2.1 and Tensorflow (>= 1.4) backend.
This repository contains a jupyter notebook and the link to the necessary data and trained models to replicate the results. Please open the notebook RCV_notebook.ipynb for more information.
Recent work on Testing with Concept Activation Vectors (TCAV) proposed directional derivatives to quantify the in uence of user-defined concepts on the network output [1]. However, diagnostic concepts are often continuous measures that might be counter intuitive to describe by their presence or absence. In this paper, we extend TCAV from a classification problem to a regression problem by computing Regression Concept Vectors (RCVs). Instead of seeking a discriminator between two concepts, we seek the direction of greatest increase of the measures for a single continuous concept. We measure the relevance of a concept with bidirectional relevance scores, Br. The Br scores assume positive values when increasing values of the concept measures positively affect classification and negative in the opposite case.
As a prior analysis, we compute the Pearson product-moment correlation coefficient between the concept measures and the network prediction for a set of patches. More information about the Correlation Analysis and the concept measures used can be found in the paper.
| correlation | ASM | eccentricity | Euler | area | contrast | |
|---|---|---|---|---|---|---|
| r | -0.2285 | -0.1869 | -0.1460 | 0.1534 | 0.2820 | 0.4119 |
| p-value | 0.001 | 0.001 | 0.01 | 0.001 | 0.001 | 0.001 |
The performance of the linear regression was computed for all the patches over multiple reruns to check if the network is learning the concepts and in which layers. The learning of the concepts across layers is linked to the size of the receptive field of the neurons and the increasing complexity of the sought patterns. Hence, more abstract concepts, potentially useful in other applications, can be learned and analyzed in deep layers of the network.
Br scores were formulated by taking into account the principles of explanation continuity and selectivity proposed in [2]. For the former, we consider whether the sensitivity scores are similar for similar data samples. For the latter, we redistribute the final relevance to concepts with the strongest impact on the decision function. More details about the definition of Br scores are reported in the paper.
We applied RCVs to eye data. See this repository for more details.
- [1] Interpretability Beyond Feature Attribution: Quantitative Testing with Concept Activation Vectors (https://arxiv.org/abs/1711.11279; Kim, B. et al https://github.com/tensorflow/tcav/blob/master/README.md)
- [2] Methods for interpreting and understanding deep neural networks. Montavon, G. et al
Clone the repository Download the model weights (tumor_classifier.h5) in /models/0528-1559/ from: https://www.dropbox.com/sh/a56pyl4i8oj9zl9/AADR74UXUu_uZA6Omb3B9Q1Ea?dl=0 For resnet101 download the pretrained imagenet weights (store in scripts/models/imagenet_weights) from: https://drive.google.com/file/d/0Byy2AcGyEVxfTmRRVmpGWDczaXM/view
Download a subsample of the preprocessed Camelyon data (store in /data/camelyon17/) from: https://github.com/medgift/PROCESS_UC1/blob/master/data/intermediate_datasets/cam17/patches.hdf5
To install the list of dependencies run:
pip install -r requirements.txt
To Run in Anaconda virtual environment (recommended):
conda create -n venv pip python=3.5
source activate venv
pip install -r requirements.txt
Note: install tensorflow-gpu for gpu otpimization
RCVs were computed by extending the keras-vis library: https://github.com/raghakot/keras-vis.
Staining normalization was performed thanks to the StainTools library: https://github.com/Peter554/StainTools.git.