Abstract:
Proper training of convolutional neural networks (CNNs) requires annotated training datasets oflarge size, which are not currently available in CT colonography (CTC). In ...Show MoreMetadata
Abstract:
Proper training of convolutional neural networks (CNNs) requires annotated training datasets oflarge size, which are not currently available in CT colonography (CTC). In this paper, we propose a well-designed framework to address the challenging problem of data shortage in the training of 3D CNN for the detection of polyp candidates, which is the first and crucial part of the computer-aided diagnosis (CAD) of CTC. Our scheme relies on the following two aspects to reduce overfitting: 1) mass data augmentation, and 2) a flat 3D residual fully convolutional network (FCN). In the first aspect, we utilize extensive rotation, translation, and scaling with continuous value to provide numerous data samples. In the second aspect, we adapt the well-known V-Net to a flat residual FCN to resolve the problem of detection other than segmentation. Our proposed framework does not rely on accurate colon segmentation nor any electrical cleansing of tagged fluid, and experimental results show that it can still achieve high sensitivity with much fewer false positives. Code has been made available at: http://github.com/chenyzstju/ctc_screening_cnn.
Published in: 2018 40th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC)
Date of Conference: 18-21 July 2018
Date Added to IEEE Xplore: 28 October 2018
ISBN Information:
ISSN Information:
PubMed ID: 30440487