Utilizing Chest X-rays for Age Prediction and Gender Classification

In this paper we present a framework for automatically predicting the gender and age of a patient using chest x-rays (CXRs). The work of this paper derives from common situations in medical imaging where the gender/age of a patient might be missing or in situations where the x-ray is of poor quality, thus leaving the medical practitioner unable to treat the patient appropriately. The proposed framework comprises of training a large CNN which jointly outputs the gender/age of a CXR. For feature extraction, transfer learning was employed using the EfficientNetB0 architecture, with a custom trainable top layer for both classification and prediction. This framework was applied to a combination of publicly available data, which collectively represent a heterogeneous dataset showing a variation in terms of race, location, patient's health, and quality of image. Our results are robust with respect to these factors, as none of them was used as input to improve the results. In conclusion, Deep Learning can be implemented in the medical imaging domain for automatically predicting characteristics of a patient.