eduzhai > Applied Sciences > Engineering >

Evaluating Knowledge Transfer in Neural Network for Medical Images

  • king
  • (0) Download
  • 20210507
  • Save

... pages left unread,continue reading

Document pages: 12 pages

Abstract: Deep learning and knowledge transfer techniques have permeated the field ofmedical imaging and are considered as key approaches for revolutionizingdiagnostic imaging practices. However, there are still challenges for thesuccessful integration of deep learning into medical imaging tasks due to alack of large annotated imaging data. To address this issue, we propose ateacher-student learning framework to transfer knowledge from a carefullypre-trained convolutional neural network (CNN) teacher to a student CNN. Inthis study, we explore the performance of knowledge transfer in the medicalimaging setting. We investigate the proposed network s performance when thestudent network is trained on a small dataset (target dataset) as well as whenteacher s and student s domains are distinct. The performances of the CNNmodels are evaluated on three medical imaging datasets including DiabeticRetinopathy, CheXpert, and ChestX-ray8. Our results indicate that theteacher-student learning framework outperforms transfer learning for smallimaging datasets. Particularly, the teacher-student learning framework improvesthe area under the ROC Curve (AUC) of the CNN model on a small sample ofCheXpert (n=5k) by 4 and on ChestX-ray8 (n=5.6k) by 9 . In addition to smalltraining data size, we also demonstrate a clear advantage of theteacher-student learning framework in the medical imaging setting compared totransfer learning. We observe that the teacher-student network holds a greatpromise not only to improve the performance of diagnosis but also to reduceoverfitting when the dataset is small.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×