Doctor of Philosophy, Tsinghua University (2019)
Bachelor of Engineering, Tsinghua University (2014)
Wireless capsule endoscopy (WCE) has developed rapidly over the last several years and now enables physicians to examine the gastrointestinal tract without surgical operation. However, a large number of images must be analyzed to obtain a diagnosis. Deep convolutional neural networks (CNNs) have demonstrated impressive performance in different computer vision tasks. Thus, in this work, we aim to explore the feasibility of deep learning for ulcer recognition and optimize a CNN-based ulcer recognition architecture for WCE images. By analyzing the ulcer recognition task and characteristics of classic deep learning networks, we propose a HAnet architecture that uses ResNet-34 as the base network and fuses hyper features from the shallow layer with deep features in deeper layers to provide final diagnostic decisions. 1,416 independent WCE videos are collected for this study. The overall test accuracy of our HAnet is 92.05%, and its sensitivity and specificity are 91.64% and 92.42%, respectively. According to our comparisons of F1, F2, and ROC-AUC, the proposed method performs better than several off-the-shelf CNN models, including VGG, DenseNet, and Inception-ResNet-v2, and classical machine learning methods with handcrafted features for WCE image classification. Overall, this study demonstrates that recognizing ulcers in WCE images via the deep CNN method is feasible and could help reduce the tedious image reading work of physicians. Moreover, our HAnet architecture tailored for this problem gives a fine choice for the design of network structure.
View details for DOI 10.1155/2019/7546215
View details for Web of Science ID 000488776000001
View details for PubMedID 31641370
View details for PubMedCentralID PMC6766681
Compared with conventional gastroscopy which is invasive and painful, wireless capsule endoscopy (WCE) can provide noninvasive examination of gastrointestinal (GI) tract. The WCE video can effectively support physicians to reach a diagnostic decision while a huge number of images need to be analyzed (more than 50 000 frames per patient). In this paper, we propose a computer-aided diagnosis method called second glance (secG) detection framework for automatic detection of ulcers based on deep convolutional neural networks that provides both classification confidence and bounding box of lesion area. We evaluated its performance on a large dataset that consists of 1504 patient cases (the largest WCE ulcer dataset to our best knowledge, 1076 cases with ulcers, 428 normal cases). We use 15 781 ulcer frames from 753 ulcer cases and 17 138 normal frames from 300 normal cases for training. Validation dataset consists of 2040 ulcer frames from 108 cases and 2319 frames from 43 normal cases. For test, we use 4917 ulcer frames from 215 ulcer cases and 5007 frames from 85 normal cases. Test results demonstrate the 0.9469 ROC-AUC of the proposed secG detection framework outperforms state-of-the-art detection frameworks including Faster-RCNN (0.9014) and SSD-300 (0.8355), which implies the effectiveness of our method. From the ulcer size analysis, we find the detection of ulcers is highly related to the size. For ulcers with size larger than 1% of the full image size, the sensitivity exceeds 92.00%. For ulcers that are smaller than 1% of the full image size, the sensitivity is around 85.00%. The overall sensitivity, specificity and accuracy are 89.71%, 90.48% and 90.10%, at a threshold value of 0.6706, which implies the potential of the proposed method to suppress oversights and to reduce the burden of physicians.
View details for DOI 10.1088/1361-6560/ab5086
View details for PubMedID 31645019
Photon-counting detectors can obtain the spectral information from an incident x-ray spectrum, although the detected counts may differ from the incident counts due to the detector response. If uncorrected or uncompensated, the response will lead to distortion in CT reconstruction. With the intention of reducing the distortion and exploring the potential of photon-counting detectors, a novel reconstruction strategy with spectral deconvolution, which attempts to set itself apart from traditional material decomposition frameworks, is proposed in this paper. It applies deconvolution to the energy window counts using a calibrated detector response and then uses the post-deconvolution photon counts to reconstruct images in multi-energy windows. The output has a quantitative meaning as a quasi-monochromatic attenuation coefficient, because a relatively narrow energy window width is selected. The deconvolution settings and results are carefully discussed in the numerical simulation. An experimental study is then carried out to verify the effectiveness and robustness. The results show that the reconstructed attenuation coefficients after deconvolution fit the standard reference data very well in most of the energy windows, which implies the feasibility of this quasi-monochromatic imaging method.
View details for DOI 10.1088/1361-6560/aa5a47
View details for Web of Science ID 000395801000004
View details for PubMedID 28099156
View details for Web of Science ID 000455836200293
View details for Web of Science ID 000432419500116