Paper Info
Title
Cxnet-M2: A Deep Model With Visual And Clinical Contexts For Image-Based Detection Of Multiple Lesions
Abstract
Diagnosing multiple lesions on images is facing with challenges of incomplete and incorrect disease detection. In this paper, we propose a deep model called CXNet-m2 for the detection of multiple lesions on chest X-ray images. In our model, there is a convolutional neural network (CNN) for encoding the images, a recurrent neural network (RNN) for generating the next word (the name of lesion) and an attention mechanism to align the visual contexts with the prediction of words. There are two main contributions of CXNet-m2 to improve the work efficiency and increase the diagnosis accuracy. (1) Inspired by image captioning, CXNet-m2 adapts the classification system to a language model, where Bi-LSTM is used to learn the clinical relationship between lesions. (2) Inspired by attention mechanism, the prediction of possible lesions is guided by visual contexts, where the visual contexts are selected by the previously generated words and chosen visual regions.The experimental results on Chestx-ray14 show that CXNet-m2 achieves better AUC and the different versions of CXNet-m2 illustrate the importance of pre-training and clinical contexts.
Year
DOI
Venue
2019
10.1007/978-3-030-23597-0_33
WIRELESS ALGORITHMS, SYSTEMS, AND APPLICATIONS, WASA 2019
Keywords
Field
DocType
Chest X-Rays image, Multi-label classification, Neural network
Closed captioning,Pattern recognition,Computer science,Convolutional neural network,Image based,Recurrent neural network,Multi-label classification,Artificial intelligence,Artificial neural network,Language model,Distributed computing,Encoding (memory)
Conference
Volume
ISSN
Citations 
11604
0302-9743
0
PageRank 
References 
Authors
0.34
0
4
Name
Order
Citations
PageRank
Shuaijing Xu172.40
Guangzhi Zhang262.25
Rongfang Bie354768.23
Anton Kos48017.96