Observational Supervision for Medical Image Classification using Gaze Data
This code implements the Gaze-WS and Gaze-MTL methods described in our paper (MICCAI 2021):
Khaled Saab, Sarah Hooper, Nimit Sohoni, Jupinder Parmar, Brian Pogatchnik, Sen Wu, Jared Dunnmon, Hongyang Zhang, Daniel Rubin, and Christopher Ré
Observational Supervision for Medical Image Classification using Gaze Data
Abstract
Deep learning models have demonstrated favorable performance on many medical image classification tasks. However, they rely on expensive handlabeled datasets that are time-consuming to create. In this work, we explore a new supervision source to training deep learning models by using gaze data that is passively and cheaply collected during a clinician’s workflow. We focus on three medical imaging tasks, including classifying chest X-ray scans for pneumothorax and brain MRI slices for metastasis, two of which we curated gaze data for. The gaze data consists of a sequence of fixation locations on the image from an expert trying to identify an abnormality. Hence, the gaze data contains rich information about the image that can be used as a powerful supervision source. We first identify a set of gaze features and show that they indeed contain classdiscriminative information. Then, we propose two methods for incorporating gaze features into deep learning pipelines. When no task labels are available, we combine multiple gaze features to extract weak labels and use them as the sole source of supervision (Gaze-WS). When task labels are available, we propose to use the gaze features as auxiliary task labels in a multi-task learning framework (Gaze-MTL). On three medical image classification tasks, our Gaze-WS method without task labels comes within 5 AUROC points (1.7 precision points) of models trained with task labels With task labels, our Gaze-MTL method can improve performance by 2.4 AUROC points (4 precision points) over multiple baselines.
Setup instructions
Prerequisites: Make sure you have Python>=3.6 and PyTorch>=1.4 installed. Then, install dependencies with:
pip install -r requirements.txtNext, either add the base directory of the repository to your PYTHONPATH, or run:
pip install -e .Configuration Options
To train a model, run the train_model.sh script.
Important configurations:
- DATA_DIR: directory where the images are saved
- SOURCE: the dataset, which is one of {cxr, cxr2, mets}, where cxr is CXR-P and cxr2 is CXR-A
- TASK: the method, which is one of {original, weak_gaze, or gaze-mtl}, where original is Image-Only and weak_gaze is Gaze-WS
Datasets
We release our gaze data on our two novel datasets for CXR-P and METS.
Images
- CXR-P: Download the dataset from the stage-1 SIIM-ACR Pneumothorax Segmentation challenge (official link, second link).
- CXR-A: Download the dataset from the "Eye Gaze Data for Chest X-rays" paper on PhysioNet.
- METS: We are currently working on the necessary PHI protocols within our hospital to make this dataset public.
Gaze data
Gaze data for the three datasets, with a visualization demo and README, can be found in the gaze_data directory.
Citation
If you use this codebase, or otherwise found our work valuable, please cite:
@inproceedings{saab2021observational,
title={Observational Supervision for Medical Image Classification Using Gaze Data},
author={Saab, Khaled and Hooper, Sarah M and Sohoni, Nimit S and Parmar, Jupinder and Pogatchnik, Brian and Wu, Sen and Dunnmon, Jared A and Zhang, Hongyang R and Rubin, Daniel and R{\'e}, Christopher},
booktitle={International Conference on Medical Image Computing and Computer-Assisted Intervention},
pages={603--614},
year={2021},
organization={Springer}
}