Can we trust deep learning models diagnosis? The impact of domain shift in chest radiograph classification

By Eduardo Henrique Pais Pooch, Pedro L. Ballester, Rodrigo C. Barros. 2019

163 citations recorded by the source corpus.

While deep learning models become more widespread, their ability to handle unseen data and generalize for any scenario is yet to be challenged. In medical imaging, there is a high heterogeneity of distributions among images based on the equipment that generates them and their parametrization. This heterogeneity triggers a common issue in machine learning called domain shift, which represents the difference between the training data distribution and the distribution of where a model is employed. A high domain shift tends to implicate in a poor generalization performance from the models. In this work, we evaluate the extent of domain shift on four of the largest datasets of chest radiographs. We show how training and testing with different datasets (e.g., training in ChestX-ray14 and testing in CheXpert) drastically affects model performance, posing a big question over the reliability of deep learning models trained on public datasets. We also show that models trained on CheXpert and MIMIC-CXR generalize better to other datasets.

Subfields: Pulmonary and Respiratory Medicine, Radiology, Nuclear Medicine and Imaging

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Pulmonary and Respiratory Medicine

1. Pranav Rajpurkar
2. Andrew Y. Ng

In Radiology, Nuclear Medicine and Imaging

1. Pranav Rajpurkar
2. Yuanda Zhu
3. Li Tong
4. Anirudh Choudhary
5. May D. Wang
6. Andrew Y. Ng

Citations in the corpus

• Application of Artificial Intelligence Technology in Oncology: Towards the Establishment of Precision Medicine (2020, methodology): “When the data were evaluated by splitting the data into 80% training (10% of which were validated) and 20% testing, it was found that the models trained with the data from the same facility as the test data were more accurate, whereas the models trained with data from a different facility were found”
• Expert-level detection of pathologies from unannotated chest X-ray images via self-supervised learning (2022, methodology): “The PadChest dataset is a public dataset that contains 160,868 chest X-ray images labelled with 174 different radiographic findings, 19 differential diagnoses 19 .”
• Domain Adaptation for Medical Image Analysis: A Survey (2021, unknown): “Howev r, many conventional mac ine l arning methods ignore this problem, which would lead to performance degradation [26], [30].”
• Fairness Violations and Mitigation under Covariate Shift (2019, background): “Examples of the impact of distribution shift can be found in medical imaging tasks [57, 75], where the algorithms trained on one chest radiography dataset performed poorly on other datasets.”
• Advancing Medical Imaging Informatics by Deep Learning-Based Domain Adaptation (2020, background): “Diversifying the training data by creating larger datasets is a possible solution, but recent medical imaging studies [21, 23, 24] have shown that it does not guarantee improved generalization.”
• Guidelines for clinical trial protocols for interventions involving artificial intelligence: the SPIRIT-AI Extension (2020, unknown): “There are limitations to the generalizability of AI algorithms, one of which is when they are used outside of their development environment 37,38 .”