Adaptive Self-Paced Transfer Learning for COVID-19 Diagnosis with CXR Images

Abstract

Coronavirus Disease 2019 (COVID-19) has become an unprecedented public health crisis since December of 2019. Compared with real-time reverse transcription polymerase chain reaction (rRT-PCR), the computer-aided diagnosis machine learning algorithm based on medical images can vastly ease the burden on clinicians. Even so, despite existing hundreds of millions of confirmed cases worldwide, there has not been a mature, large scale, high quality, single standard shared image data set yet, which can lead to some problems. For instance, 1) Because the sources of medical images and the collection standards are not guaranteed, features extracted by the neural network may not be very ideal. 2) Due to the small number of samples, some outliers (e.g., blurry medical images, inconspicuous symptoms) may significantly descend the performance of the model. To address these problems, we propose an adaptive self-paced transfer learning (ASPTL) algorithm in this paper. Specifically, inspired by the process of human learning from easy to difficult, we also evaluated the learning difficulty of the samples. Samples with no obvious disease features or wrong labels are relatively difficult to diagnose, and the samples that are easy to diagnose are selected adaptively in the iterative process. In addition, we adopt transfer learning to select easy to learn samples on the pre-trained network by self-paced learning, and gradually fine-tune the pre-trained model in an iterative way. We designed two experiments to validate the ASPTL algorithm’s performance on COVID-19. The reult prove the effectiveness on solving mentioned problems.