Data-efficient deep learning algorithms for computer-aided medical diagnosis

Abstract

As a critical component of many healthcare applications, such as diagnosis and surgical planning, precise and robust segmentation of organs and lesions from medical images is crucial. Deep learning has been successful in general image segmentation due to the increasing amount of annotation data available. However, the acquisition of labeled data for medical images is usually expensive since generating accurate annotations requires professional knowledge and time. Therefore, in this research, we have outlined three specific objectives to achieve data-efficient deep learning algorithms for computer-aided medical diagnosis: (1) synthesizing raw data to enhance the performance of deep learning algorithms and solve the issue of medical data shortage and unbalanced class; (2) maximizing the performance of supervised deep learning algorithms by designing advanced architecture; (3) utilizing unannotated medical images to address the problem of scarcity of annotated medical image data; Our experiments indicate that our proposed deep learning algorithms outperform other state-of-the-art models. Further research needs to be conducted to determine the feasibility and reliability of applying deep learning models to real clinical applications. The research in medical image segmentation has the potential to enable the implementation of automatic disease diagnosis and surgical planning in real clinical scenarios.