Document Type

Conference Proceeding

Publication Date



Transfer Learning methods are extensively applied with standard CNN architectures for various medical diagnoses. However, these architectures are computationally expensive, tend to be over parameterized, and requires a relatively large labeled datasets which are often not available in the medical image domain. Accordingly, this paper proposes a Multi-Stage Transfer Learning System using lightweight architectures to address problems with limited data and to improve training time. Preliminary results suggest that our model performed well on CT Head images over traditional single-stage transfer learning.