Eliminating race-related shortcuts in deep neural networks for chest X-ray analysis
Published in AAAI 2022 Workshop: Trustworthy AI for Healthcare, 2022
We applied several image processing methods to identified the key information in chest X-ray that CNN-based deep learning model used for classifying patient’s race group. Our results revealed that the shape of lungs is an important factor in the classification of race and is therefore a likely shortcut for the detection of radiological features. The image rotation can be used to decrease the weight assigned to race-related features without compromising the performance of anomaly detection. The proposed training scheme was also shown to mediate the disparities in detection performance among races.
Recommended citation: Wang, R., Chen, L.C., Lin, P.C., Wawira, J., Celi, L., & Kuo, P.C. (2022). Eliminating race-related shortcuts in deep neural networks for chest X-ray analysis. AAAI 2022 Workshop: Trustworthy AI for Healthcare.
Download Paper