Advanced Techniques in Chest X-ray Image Recognition

Chest Disease Recognition from X-ray Image Group 2 JunyangYao Yijun Yan Ziling Yuan UC San Diego June 3,2020 Group 2, Junyang Yao, Yijun Yan, Ziling Yuan Chest Disease Recognition from X-rayImage 1/ 19

Background The issue of recognizing X-rays images has existed for many years. A 1999 study found that major disagreement between 2 observers in interpreting x-rays of patients in an emergency department happened in 5-9% of total cases. Recognizing by naked eye is not reliable based on our human s biological properties, which is largely influenced by doctors experience or otherfactors. Group 2, Junyang Yao, Yijun Yan, Ziling Yuan Chest Disease Recognition from X-rayImage 2/ 19

Background Therefore, the motivation is based on such facts to build a deep convolutional neural network (CNN) and depthwise separable convolution system can helpthe frontline give Chest X-ray (CXR) reports more accurate and faster. Figure: Chest X-ray image Group 2, Junyang Yao, Yijun Yan, Ziling Yuan Chest Disease Recognition from X-rayImage 3/ 19

Literature Survey From our literature survey, we attempt three approaches to solve theproblem. MobileNets Andrew G. Howard et al present a class ofefficient models called MobileNets for mobile and embedded vision applications [1]. The MobileNets model is based on depthwise separable convolutions which is a form ofconvolutions which factorize a standard convolution into a depthwise convolution and a 1x1 convolution called a pointwise convolution [1]. In MobileNets these depthwise convolution layers are like filters to each channel, so only useful information arekept. Group 2, Junyang Yao, Yijun Yan, Ziling Yuan Chest Disease Recognition from X-rayImage 4/ 19

Literature Survey Xception In 2016, Fran ois Chollet presents an interpretation of Inception modulesin convolutional neural networks as being an intermediate step in-between regular convolution and the depthwise separable convolution operation[2]. Xception is a convolutional neural network that is 71 layers deep, which is based on the observation that a depthwise separable convolution can be understood as an Inception module with a maximally large number of towers. Group 2, Junyang Yao, Yijun Yan, Ziling Yuan Chest Disease Recognition from X-rayImage 5/ 19

Literature Survey ConvNet Karen Simonyan and Andrew Zisserman investigate the effect of convolutional network depth on the accuracy in the large-scale image recognition[3]. They use some configuration make some change to CNN architecture. Group 2, Junyang Yao, Yijun Yan, Ziling Yuan Chest Disease Recognition from X-rayImage 6/ 19

Dataset Dataset source: https://data.mendeley.com/datasets/rscbjbr9sj/2 Divided into 3 folders: test, train, val (a) TestingSet (c) Validation Set (b) TrainingSet Figure: Size of dataset Group 2, Junyang Yao, Yijun Yan, Ziling Yuan Chest Disease Recognition from X-rayImage 7/ 19

Feature Extraction Pixels (after resizing) ! Group 2, Junyang Yao, Yijun Yan, Ziling Yuan Chest Disease Recognition from X-rayImage 8/ 19

Models CNN-based Architecture Model: MobileNet Xception-based Model Proposed ConvNet-basedModel Group 2, Junyang Yao, Yijun Yan, Ziling Yuan Chest Disease Recognition from X-rayImage 9/ 19

MobileNet Group 2, Junyang Yao, Yijun Yan, Ziling Yuan Chest Disease Recognition from X-rayImage 10/ 19

Xception Group 2, Junyang Yao, Yijun Yan, Ziling Yuan Chest Disease Recognition from X-rayImage 11/ 19

Simplified Xception Remove the red box. Group 2, Junyang Yao, Yijun Yan, Ziling Yuan Chest Disease Recognition from X-rayImage 12/ 19

Proposed ConvNet-based Model Group 2, Junyang Yao, Yijun Yan, Ziling Yuan Chest Disease Recognition from X-rayImage 13/ 19

Proposed ConvNet-based Model Optimizations we made: 1. Use batch-normalization with convolutions to avoidoverfitting. 2. Replace convolutional layers with depthwise separable convolutionallayers. 3. Initialize the first few layers from a network pretrained on ImageNet instead of randomly initialized weights to have a much better initialization Group 2, Junyang Yao, Yijun Yan, Ziling Yuan Chest Disease Recognition from X-rayImage 14/ 19

Results TP 388 323 386 TN 71 161 104 130 FP 163 73 FN 2 67 4 recall 0.99 0.83 0.99 precision 0.70 0.82 0.75 accuracy 73.6% 77.6% 83.5% MobileNets Xception ConvNet Table: Data of confusionmatrices. Group 2, Junyang Yao, Yijun Yan, Ziling Yuan Chest Disease Recognition from X-rayImage 15/ 19

Discussion ConvNet performs best between the threemodels. Recall is an important matrix. False positive should be as low as possible. Fine-tuning improves the model alot. A good initialization reduces much running time. Group 2, Junyang Yao, Yijun Yan, Ziling Yuan Chest Disease Recognition from X-rayImage 16/ 19

Future Work Train more epochs and make the model morestable Further fine-tuning the model and improves itsaccuracy and recall Have a trial of more complex CNNarchitectures Group 2, Junyang Yao, Yijun Yan, Ziling Yuan Chest Disease Recognition from X-rayImage 17/ 19

References Howard, Andrew G., et al. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprintarXiv:1704.04861 (2017). Chollet, Fran ois. Xception: Deep learning with depthwise separable convolutions. Proceedings of the IEEE conference on computer vision and pattern recognition. 2017. Simonyan, Karen, and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprintarXiv:1409.1556 (2014). Group 2, Junyang Yao, Yijun Yan, Ziling Yuan Chest Disease Recognition from X-rayImage 18/ 19

Thank you! Group 2, Junyang Yao, Yijun Yan, Ziling Yuan Chest Disease Recognition from X-rayImage 19/ 19