Differences between human and machine perception in medical diagnosis

This repository accompanies our paper Differences between human and machine perception in medical diagnosis. In the paper, we propose a framework for comparing human and machine perception in medical diagnosis, and demonstrate it with a case study in breast cancer screening. This repository contains the data and code necessary to reproduce the results from our case study. There are three components:

1. probabilistic_inference.py: We collected predictions from radiologists and DNNs on screening mammograms perturbed with Gaussian low-pass filtering (Figure 1a--b). The predictions are provided in data/observed_predictions, see below for details. We apply probabilistic modeling to these predictions in order to isolate the effect that low-pass filtering has on their predictions (Figure 1d).
2. perturbation_study_analysis.py: We sample from the probabilistic model and compare radiologists and DNNs with respect to predictive confidence and class separability (Figure 1e--f).
3. annotation_study_analysis.py: Radiologists annotated regions of interest (ROIs) indicating the most suspicious regions of images (Figure 1g). We then collected predictions from DNNs where we separately applied low-pass filtering to the ROI interior (Figure 1h), ROI exterior (Figure 1i), and to the entire image (Figure 1j). These predictions are provided in data/observed_predictions, see below for details. We then examine how these low-pass filtering schemes affect the DNNs' class separability.

Additionally, we include our implementation of Gaussian low-pass filtering in fourier_filter.py.

Setup

Add the directory path of this repository to PYTHONPATH, and install the following:

• Python 3.*
• Gin-config
• Imageio
• Matplotlib
• NumPy
• Pandas
• PyStan
• SciPy

PyStan requires a C++14 compatible compiler. If gcc-related errors are encountered, see the detailed installation instructions in https://pystan.readthedocs.io/en/latest/installation_beginner.html. These errors can likely be resolved with:

• Linux: conda install gcc_linux-64 gxx_linux-64 -c anaconda
• Mac: conda install clang_osx-64 clangxx_osx-64 -c anaconda

We recommend using anaconda to install everything except Gin-config and PyStan, which can be installed with pip. The expected install time on a standard desktop computer is approximately five minutes. This code has been tested on Ubuntu 16.04 and macOS Mojave 10.14.6.

Data

The radiologist and DNN predictions from our two reader studies are stored as NumPy arrays, and are located in data/observed_predictions. For DNNs, we include predictions for two architectures: GMIC and DMV.

For our perturbation study, we have predictions for (i) radiologists (radiologists.pkl), (ii) DNNs trained on unperturbed data (unperturbed.pkl), and (iii) DNNs trained on low-pass filtered data (filtered.pkl). The radiologists' predictions have the shape (10, 9, 720, 2), where 10 is the number of radiologists, 9 is the number of filtering severities, 720 is the number of screening mammography exams, and 2 is the number of breasts. The DNNs' predictions have a similar shape of (5, 9, 720, 2), where 5 is the number of DNN training seeds, and all other dimensions are the same as the radiologists. Each set of predictions comes with a corresponding mask, which is a binary NumPy array with the same shape as the predictions. Each element in the mask is set to 1 if we have a prediction for that corresponding index. This is necessary since the radiologists' predictions are sparse, while the DNNs' predictions are dense. See our paper for details regarding this issue.

For our annotation study, we have DNN predictions where low-pass filtering is applied to the (i) ROI interior (roi_interior.pkl), (ii) ROI exterior (roi_exterior.pkl), and to the (iii) entire image (these are a subset of predictions from the perturbation study). These predictions have the shape (7, 5, 9, 120, 2), where 7 is the number of radiologists who annotated the images, 5 is the number of DNN training seeds, 9 is the number of severities, 120 is the number of screening mammography exams, and 2 is the number of breasts.

Usage

Reproducing our results

Here are the steps for reproducing our results on comparing radiologists and DNNs for the GMIC architecture. See demo.ipynb for these steps applied sequentially. To perform the analyses with the DMV architecture, replace all instances of gmic with dmv. The results of the analyses are saved in the results directory by default, but this behavior can be modified in the configuration files in cfg.

The first step is to perform probabilistic inference and generate posterior samples. The outputs are NumPy arrays, which will be used in the subsequent analysis.

1. Radiologists: python code/probabilistic_inference.py cfg/probabilistic_inference/radiologists.gin
2. DNNs: python code/probabilistic_inference.py cfg/probabilistic_inference/dnns/gmic/unperturbed.gin
3. DNNs (trained w/ filtered data): python code/probabilistic_inference.py cfg/probabilistic_inference/dnns/gmic/filtered.gin

Next, we compare how low-pass filtering affects the predictive confidence and class separability of radiologists and DNNs, performing separate analyses for two subgroups. The outputs are pdf images which correspond to Figure 4 in the paper.

1. Microcalcifications: python code/perturbation_study_analysis.py cfg/perturbation_study_analysis/gmic/microcalcifications.gin
2. Soft tissue lesions: python code/perturbation_study_analysis.py cfg/perturbation_study_analysis/gmic/soft_tissue_lesions.gin

Finally, we compare radiologists and DNNs with respect to the regions of an image deemed most suspicious. We perform separate analyses for two subgroups, but this time in a single call. The output is a pdf image corresponding to Figure 5 in the paper.

python annotation_study_analysis.py cfg/annotation_study_analysis/gmic.gin

The combined expected run time on a standard desktop computer is approximately 20 minutes per model architecture.

Gaussian low-pass filtering

See fourier_filter.ipynb to see our implementation of Gaussian low-pass filtering applied to an example screening mammogram.

License

This repository is licensed under the terms of the GNU AGPLv3 license.

Reference

If you found this code useful, please cite our paper:

Differences between human and machine perception in medical diagnosis
Taro Makino, Stanisław Jastrzębski, Witold Oleszkiewicz, Celin Chacko, Robin Ehrenpreis, Naziya Samreen, Chloe Chhor, Eric Kim, Jiyon Lee, Kristine Pysarenko, Beatriu Reig, Hildegard Toth, Divya Awal, Linda Du, Alice Kim, James Park, Daniel K. Sodickson, Laura Heacock, Linda Moy, Kyunghyun Cho, Krzysztof J. Geras
2020

@article{Makino2020Differences, 
    title = {Differences between human and machine perception in medical diagnosis},
    author = {Taro Makino and Stanisław Jastrzębski and Witold Oleszkiewicz and Celin Chacko and Robin Ehrenpreis and Naziya Samreen and Chloe Chhor and Eric Kim and Jiyon Lee and Kristine Pysarenko and Beatriu Reig and Hildegard Toth and Divya Awal and Linda Du and Alice Kim and James Park and Daniel K. Sodickson and Laura Heacock and Linda Moy and Kyunghyun Cho and Krzysztof J. Geras}, 
    journal = {arXiv:2011.14036},
    year = {2020}
}