Neural Prototype Tree for Chest X-Ray Pathology Detection
Author Information
Author(s): Mavragani Amaryllis, Zhang Haoran, Arasteh Soroosh Tayebi, Chen Hongbo, Alfred Myrtede, Brown Andrew D, Atinga Angela, Cohen Eldan
Primary Institution: University of Toronto
Hypothesis
This study aimed to investigate the utility of the NPT classifier in performance, interpretability, and fairness.
Conclusion
The NPT classifier can achieve competitive performance while providing interpretable explanations for its decision-making process.
Supporting Evidence
- The NPT classifier's performance improved with higher interpretability levels.
- The NPT classifier surpassed ResNet-152 in performance at certain interpretability levels.
- A significant positive relationship was found between interpretability and performance.
- The mean TPR disparity indicated varying levels of unfairness across demographic groups.
Takeaway
This study created a new tool that helps doctors understand chest X-ray results better while also being fair to all patients.
Methodology
The study used CXRs from publicly available datasets and trained multiple classifiers to evaluate performance, interpretability, and fairness.
Potential Biases
The NPT classifiers exhibited biases across demographic groups, particularly in age and sex.
Limitations
The datasets were automatically labeled, which may lead to inaccuracies, and the study did not include all available datasets due to resource constraints.
Participant Demographics
{"Chest X-ray 14":{"Female":"13,403 (43.51%)","Male":"17,402 (56.49%)"},"CheXpert":{"Female":"29,019 (44.48%)","Male":"36,221 (55.52%)"},"MIMIC-CXR":{"Female":"34,252 (52.39%)","Male":"31,127 (47.61%)"}}
Statistical Information
P-Value
p<0.001
Statistical Significance
p<0.001
Digital Object Identifier (DOI)
Want to read the original?
Access the complete publication on the publisher's website