代写 C algorithm deep learning Scheme html parallel database graph statistic network This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JBHI.2019.2922682, IEEE Journal of Biomedical and Health Informatics

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JBHI.2019.2922682, IEEE Journal of Biomedical and Health Informatics
JBHI-01179-2018
been achieved [13]. However, existing analysis toward cervigrams only focused on a single frame, which shows the most obvious changes in the acetic-white test, ignoring the fact that the test results are determined by the dynamic characteristics of the lesions. Furthermore, existing studies have not considered the other two states of images, which contribute greatly to the doctor’s judgment. The clinical effectiveness of such methods is limited by the input data.
The current paper presents the construction of C-RCNN to address sequential and multistate cervigrams. C-RCNN is a computer-assisted algorithm that can be applied in routine colposcopy and help physicians for auxiliary diagnosis. We leverage not only CNN models to consider the spatial features of each frame but also Long Short-Term Memory (LSTM) to incorporate the extraction of temporal features into the overall architecture. Moreover, we integrate features of three different cervigram states together and train the network to classify normal, CIN1, CIN2/3, and cervical cancer cervigrams. The experimental results are compared with those of some competing classifiers on the same dataset, whereas test accuracy, sensitivity, specificity, missed diagnosis rate, misdiagnosis rate, and AUC are tested to evaluate performance.
The main contributions of this work are summarized as follows:
1) A novel CIN grade and cervical cancer classification method is presented to analyze sequential and multistate cervigram images automatically. In contrast with alternative methods, C-RCNN can effectively encode spatio-temporal information and extract high-level representations in a data- driven way.
2) To obtain dynamic information of lesions precisely, we designed a sequence encoding module based on LSTM and a constructed voting layer that can produce more discriminative temporal features than previously.
3) To integrate multistate features in the overall architecture, a concatenate layer and a multistate-aware convolutional layer are presented for integrating and realizing dimensionality reduction for features from different states.
4) Our proposed method is evaluated with 4,753 real cervigram images. Our achieved results outperform other approaches by a significant margin. The overall test accuracy reaches 96.13% with a specificity and sensitivity of 98.22% and 95.09%, respectively. The AUC is above 0.94 for each ROC curve.
The remainder of this paper is organized as follows. Section 2 provides an overview on the related work for computer- assisted methods, especially deep learning algorithms for cervigram image analysis. Section 3 provides detailed explanations of our proposed C-RCNN. Experimental results, evaluation metrics, and comparisons with other methods are discussed in Section 4. Section 5 concludes the paper with future research directions.
2
classification performance of cervigrams, and can be a potentially widely accessible automatic screening method for cervical cancer. Over the past decade, previous studies [14], [15] employed hand-crafted features to represent images and combined them with other examination information, including PAP/HPV test results and patient’s past medical history, to calculate classification probabilities. Instead of only considering color features, adding texture and other varied types of hand-crafted features can increase the accuracy in recognizing important vascular patterns [16]. A filter bank of texture models has been used for recognizing punctuation and mosaicism, resulted in a positioning accuracy exceeding 95%. However, the selection of hand-crafted features requires intensive professional knowledge, and manually designing proper features that are fusible across different modalities is difficult [13]. By contrast, neural networks, SVM, KNN, linear discriminant analysis, and decision trees can be applied to analyze cervigram images automatically [17]. For example, linear SVM and KNN have been employed to classify cervigrams, and the accuracy obtained using a neural network is higher than that of traditional methods [13]. Deep learning methods to automatically determine the characteristics of lesions exhibit improved performance [12], [13]. The use of deep learning methods in medical image recognition is a current research trend [18]–[24].
B. Sequential Medical Image Analysis
Similar to cervigram images, medical images, such as endoscopic videos and MRI pictures, are composed of sequential images. The temporal features contained in sequential medical images have been analyzed. For the segmentation of brain images [25], LSTM has been used instead of 3D CNN for MR image analysis, achieving better results than existing methods. The complementary information of visual and temporal features learned from a CNN with LSTM has been utilized, producing additional discriminative spatio- temporal features to largely boost the surgical workflow recognition problem [26]. A deep learning framework combines convolution layers, deconvolution layers, and LSTM layer for identifying and classifying breast cancer images [27]. These methods inspired us to use CNN with LSTM for analyzing sequential images. This paper presents the construction of C-RCNN, which extracts the spatio-temporal features of cervigram images to improve the classification accuracy of each CIN grade and cervical cancer.
III. METHODOLOGY
A. Overall
In this section, the proposed computer-assisted algorithm is introduced briefly. Our proposed C-RCNN is an essential part of a clinical decision support system for cervical cancer screening, which should also contain the cervigram acquisition module to shoot multistate image sequences, as well as the output module to display the classification results to physicians for auxiliary diagnosis. An overview of C-RCNN is shown in Fig. 1. Based on the multistate and sequential images obtained
A.
II. RELATED WORK
Computer-Assisted Algorithms for Cervigrams Analysis
Computer-assisted algorithms can effectively improve the
2168-2194 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JBHI.2019.2922682, IEEE Journal of Biomedical and Health Informatics
JBHI-01179-2018
Finally, the fully connected layer reduces all of the feature maps into a one-dimensional spatial feature vector 𝑦.
For the sequential images of the acetic-white test, the images are input into the CNN model frame by frame to ensure that the spatial features of each frame can be extracted directly. The generated vector sequence {𝑦1 …𝑦𝑛} is then passed to the sequence encoding module. For the images photographed after the green lens and compound iodine solution are used, only their discriminative spatial features are considered, and the feature vector {𝑧2, 𝑧3} is passed to the concatenate layer, which is constructed before the multistate-aware convolutional layer.
C. Sequence Encoding Module Based on LSTM
Given that the acetic-white test is a dynamic process in that the color of abnormal cells gradually turns white, the discrimination of analyzing a single image can only obtain a static appearance of the spatial information, and the CIN grades cannot be inferred from the characteristics of the lesions contained in only one frame. Therefore, the analysis of the acetic-white test must account for the dependency relationships between the successive frames and extract sequential dynamic features frame by frame, which will improve the discrimination accuracy of the overall result.
Compared with the traditional RNN, LSTM can learn long- term dependencies, and the gradient does not tend to vanish when trained with back propagation through time due to its special architecture [31]. The combination of CNN and LSTM achieves good results in video caption generation and target recognition [32], [33], which can also be applied to medical image analysis because it can discover more discriminative spatio-temporal features [26], [34] than previously. Thus, LSTM is used according to our sequential-level images.
In this section, we design a sequence-encoding module based on LSTM. The architecture of this module consists of an LSTM layer, a voting layer, and a fully connected layer, which has the input sequence {𝑦1 … 𝑦𝑛} generated by the CNN. After extracting temporal features, the spatio-temporal information can be well encoded in the output vector 𝑧1. The LSTM layer employs 256 LSTM units and dropout is adopted in our encoding module to prevent overfitting.
4
the vanishing gradient problem. The calculation formulas of LSTM are as follows:
𝑖 =𝜎(𝑊 𝑟 +𝑊 h +𝑏), 𝑡 𝑟𝑖𝑡h𝑖𝑡−1𝑖
𝑓=𝜎(𝑊 𝑟+𝑊 h +𝑏), 𝑡 𝑟𝑓𝑡h𝑓𝑡−1𝑓
𝑜 = 𝜎(𝑊 𝑟 + 𝑊 h + 𝑏 ), 𝑡 𝑟𝑜𝑡 h𝑜𝑡−1 𝑜
𝑔 = 𝜑(𝑊 𝑟 + 𝑊 h + 𝑏 ), 𝑡 𝑟𝑔𝑡𝑔h𝑡−1𝑔
𝑐 =𝑓𝑐 +𝑖𝑔, 𝑡 𝑡 𝑡−1 𝑡 𝑡
h𝑡 = 𝑜𝑡φ(𝑐𝑡),
For the input sequence {𝑦1 … 𝑦𝑛 }, 𝑦𝑡 represents the input at
time step t, and the output h𝑡−1 of the previous moment is used to calculate the current unit states. The LSTM units account for the previous state during the calculation of the current gates, where 𝜎 represents the hard-sigmoid nonlinear activation
function 𝜎(𝑎) = 1 to normalize the values, φ is the Tanh 1+𝑒−𝑎
nonlinear activation function 𝜑(𝑎) = 𝑒𝑎−𝑒−𝑎 that maps real 𝑒𝑎+𝑒−𝑎
values to (-1, 1),  is an elementwise multiplication that involves computations with gates, and the sets {W} and {b} represent the weight matrix and bias, respectively.
For the output sequence {h1, h2, … , h𝑛} of the LSTM layer, existing research works always directly treat the last hidden state h𝑛 as the final descriptor. However, for the problem of cervigram classification, previous states also contain valuable information for doctors to make diagnoses. Thus, we averaged all state’s output to produce additional discriminative results and improve classification accuracy by our constructed voting layer.
𝑧=(∑𝑛 h)/𝑛 (2) 1 𝑡=1𝑡
Therefore, for the output vector sequence {h1, h2, … , h𝑛} of the LSTM layer, we integrate and reduce its dimensionality to the final spatio-temporal feature 𝑧1 by the voting layer and a constructed fully connected layer.
D. Multistate-Aware Convolutional Layer for CIN Grade Prediction
After the above two steps, the CNN-LSTM part of the C- RCNN architecture has fully extracted the spatio-temporal features within each state; however, to generate the colposcopy results, the differences between the successive states [9], [10] also affect the final diagnosis, which must be considered. The characteristics of each state must be considered to make a final prediction.
For this reason, a concatenate layer is leveraged to integrate feature vectors {𝑧1, 𝑧2, 𝑧3} to a feature matrix Z. Moreover, a multistate-aware convolutional layer is constructed to reduce the dimensionality and extract multi-state features of Z. This constructed convolutional layer is used to learn the differences between different state features and to refine state-level
TABLE I
SPECIFIC INFORMATION OF THE PROPOSED CNN ARCHITECTURE FOR SPATIAL FEATURE EXTRACTION
Layers Conv1 Pool1 Conv2 Pool2 Conv3 Conv4 Con5 Pool3 Fc1 Fc2
Kernel 7*7 3*3 5*5 3*3 3*3 3*3 3*3 3*3 – – Stride 4 2 1 2 1 1 1 2 – – Channel 96 96 128 128 256 256 128 128 512 128
The LSTM layer is composed of a series of LSTM units. Each unit employs an input gate 𝑖 , a forget gate 𝑓 , an output
𝑡𝑡
gate 𝑜𝑡, and a memory cell 𝑐𝑡. Here, 𝑖𝑡 controls how much new
information enters the unit and alters the state of the memory cell. The variable 𝑓 controls what to be remembered and what
𝑡
to be forgotten, 𝑐𝑡 is a summation of the incoming information,
and 𝑜𝑡 allows the state of the memory cell to affect the current hidden state or other units. These unique structures enable the LSTM to capture long-term temporal dynamics and overcome
2168-2194 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JBHI.2019.2922682, IEEE Journal of Biomedical and Health Informatics
JBHI-01179-2018 9
TABLEV
COMPARISON OF THE PROPOSED CNN ARCHITECTURE WITH OTHER CNNS
AlexNet
GoogLe Net
ResNet
C-RCNN
Accuracy Sp
86.75% 89.96% 90.44% 92.37% 93.21% 96.27% 96.13% 98.22%
Se 
83.40% 10.04% 87.63% 7.63% 91.88% 3.73% 95.09% 1.78%

16.60% 12.37% 8.12% 4.91%
TABLE VI
PERFORMANCE COMPARISON BETWEEN THE CNN-LSTM ARCHITECTURE AND THE CNN OR LSTM MODELS
CNN-Only LSTM-Only CNN-LSTM
Accuracy Sp
85.76% 91.18% 74.55% 86.41% 94.21% 97.76%
Se
84.06% 70.19% 93.63%

8.82%
13.59 %
2.34%

15.94% 29.81% 6.37%
image sequences, and 2D-CNN is selected to reduce the dimensionality of the multistate features in the end. In addition, we attempt to use CNN to classify different CIN grades and cervical cancer directly without considering the characteristics of the lesions, which differ from state to state. The same dataset is used to train each classifier, and 5-fold cross validation is adopted to evaluate their performances. As shown in Table IV, the test accuracy is 86.45% and 91.90%. The accuracy curves between our method and the competing approaches are shown in Fig. 5.
The experimental results show that the performance of existing methods appear to be ineffective when considering the images photographed after applying the green lens and compound iodine solution. Such methods are not applicable to analyzing multistate cervigram images, which can improve the performance actually. Existing methods cannot address sequential images and do not account for the dynamic characteristics of the lesions, except for 3D-CNN. We can conclude that the competing methods do not apply to our dataset and that they are unsuitable for clinical examination.
3) Effect of CNN parameters for spatial feature extraction:
One of the most important tasks in C-RCNN is to extract discriminative spatial features. Although increasing the number of network layers can improve the CNNs’ performance, it will also considerably require prolonged training time and occupy additional computing resources. In our experiments, several parameters of the classical CNNs’ architectures, including AlexNet, GoogLeNet, and ResNet, are adjusted to extract additional discriminative features. Finally, the obtained CNN architecture based on the AlexNet meets our requirements.
We compare the performances of different architectures that are constructed with some classical CNN models. Note that only the CNN models utilized are different but exhibits the same LSTM and following layers. Cross validation is adopted, whereas each metric is averaged. The results are shown in Table
V. Our constructed C-RCNN exhibits a test accuracy of 96.13% with a specificity and sensitivity of 98.22% and 95.09%, respectively, which is clearly superior to others. For a detailed comparison at different operating points, we also perform ROC analysis for some classical CNN architectures and the proposed C-RCNN. The comparisons are conducted for each of the considered classes, and the analysis is performed using a one- versus-all scheme. As shown in Fig. 6, the proposed C-RCNN achieves the highest AUC on each of the four classes whereas the architecture with Resnet shows a competitive performance. The performances of AlexNet and GoogLeNet are not as ideal as we expect, especially on the class of CIN1. The results of the analysis confirm that our constructed architecture show remarkably superior performance against other methods and is suitable for the problem of cervical cancer screening.
4) Effect of each part of CNN-LSTM: To evaluate each part of the CNN-LSTM network’s contribution to the overall result, we remove the CNN model and the LSTM model separately. As shown in Table VI and Fig. 7, for the part of CNN-only, we remove the LSTM layer and substitute it with a fully connected layer. The feature vector is directly classified by the Softmax layer with a test accuracy of 85.76%. For the part of LSTM- only, we construct a deep LSTM model that can capture a high level of sequence information by stacking LSTM layers. Each layer in the LSTM is a hierarchy that receives the hidden state of the previous layer as input. We utilize the entire image into the LSTM model directly and construct a fully connected layer to complete the classification, which obtains a test accuracy of 74.55%. By contrast, the application of CNN-LSTM is superior in all evaluation metrics, from which it can be inferred that using the CNN-LSTM architecture can improve the performance better than using the CNN or LSTM model separately. Considering not only the static performance but also the dynamic characteristics of the lesions is helpful in the CIN grade and cervical cancer classification.
V. CONCLUSIONS AND FUTURE WORK
In this paper, focusing on the problem of the low specificity of colposcopy examination, we propose a novel computer- assisted algorithm that is different from designing hand-crafted features to describe the visual appearance and dynamic changes or traditional machine learning methods. The proposed discriminative learning model, C-RCNN, exhibits a key novelty in using CNN to extract the spatial features while leveraging LSTM to extract the temporal features. We also consider multistate cervigrams generated in clinical examination and then integrate them all to reduce the dimension. In this manner, not only the visual representations and sequential dynamics can be jointly and effectively optimized in the training phase, but also differences between different state features can be learned to refine high-level semantic information. Compared with the competing methods, the C-RCNN shows improved performance in terms of the specificity, sensitivity, missed diagnosis rate, misdiagnosis rate, test accuracy, and AUC. Importantly, the proposed C-RCNN is a computer-assisted algorithm that can be applied into the colposcopy routine because its input data is the clinical image sequences generated
2168-2194 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JBHI.2019.2922682, IEEE Journal of Biomedical and Health Informatics
JBHI-01179-2018
during colposcopy. In clinical practice, C-RCNN is a core part of the clinical decision support system that also requires cervigram acquisition module to obtain pictures of the cervix, as well as an output module to display the classification results to physicians for auxiliary diagnosis.
In our future work, we plan to leverage different variations of RNN, such as GRU and bidirectional LSTM. We also plan to achieve object recognition tasks toward our cervigram dataset. Given that our proposed method is mainly used for the classification of different CIN grades and cervical cancer, no specific outputs of the lesion positions or lesion types can provide doctors with intuitive judgments. Therefore, we plan to try YOLO, Faster R-CNN, and other object recognition methods. We will extend this architecture to other applications in medical image analysis, such as gastroscopy and capsule gastroscopy.
[13] [14] [15]
[16]
[17] [18] [19]
[20]
[21] [22] [23]
[24]
[25]
[26] [27]
[28] [29] [30] [31] [32] [33] [34]
[35]
10
T. Xu, H. Zhang, X. Huang, S. Zhang, and D. N. Metaxas, “Multimodal deep learning for cervical dysplasia diagnosis,” in MICCAI, 2016, pp. 115–123.
D. Song et al., “Multimodal entity coreference for cervical dysplasia diagnosis,” IEEE Trans. Med. Imaging, vol. 34, no. 1, pp. 229–245, 2015.
T. Xu, X. Huang, E. Kim, L. R. Long, and S. Antani, “Multi-test cervical cancer diagnosis with missing data estimation,” in SPIE Medical Imaging, 2015, vol. 56, p. 94140X.
S. Gordon, G. Zimmerman, and H. Greenspan, “Image segmentation of uterine cervix images for indexing in PACS,” in Proceedings. 17th IEEE Symposium on Computer-Based Medical Systems, 2004, no. May, p. 298.
Y. Jusman, S. C. Ng, and N. A. Abu Osman, “Intelligent screening systems for cervical cancer,” Sci. World J., vol. 2014, 2014.
G. Litjens et al., “A Survey on Deep Learning in Medical Image Analysis,” Med. Image Anal., vol. 42, pp. 60–88, 2017.
R. Zhang et al., “Automatic Detection and Classification of Colorectal Polyps by Transferring Low-Level CNN Features from Nonmedical Domain,” IEEE J. Biomed. Heal. Informatics, vol. 21, no. 1, pp. 41–47, 2017.
J. X. Qiu, H. J. Yoon, P. A. Fearn, and G. D. Tourassi, “Deep Learning for Automated Extraction of Primary Sites from Cancer Pathology Reports,” IEEE J. Biomed. Heal. Informatics, vol. 22, no. 1, pp. 244–251, 2018.
A. Esteva et al., “Dermatologist-level classification of skin cancer with deep neural networks,” Nature, vol. 542, no. 7639, pp. 115– 118, 2017.
Z. Yan et al., “Multi-Instance Deep Learning: Discover Discriminative Local Anatomies for Bodypart Recognition,” IEEE Trans. Med. Imaging, vol. 35, no. 5, pp. 1332–1343, 2016.
V. Gulshan et al., “Development and validation of a deep learning algorithm for detection of diabetic retinopathy in retinal fundus photographs,” JAMA – J. Am. Med. Assoc., vol. 316, no. 22, pp. 2402–2410, 2016.
H. Wang, S. Ding, D. Wu, Y. Zhang, and S. Yang, “Smart connected electronic gastroscope system for gastric cancer screening using multi-column convolutional neural networks,” Int. J. Prod. Res., vol. 7543, pp. 1–12, 2018.
M. F. Stollenga, W. Byeon, M. Liwicki, and J. Schmidhuber, “Parallel Multi-Dimensional LSTM, With Application to Fast Biomedical Volumetric Image Segmentation,” Comput. Sci., pp. 1– 9, 2015.
Y. Jin et al., “SV-RCNet: Workflow Recognition from Surgical Videos using Recurrent Convolutional Network,” IEEE Trans. Med. Imaging, vol. 37, no. 5, pp. 1114–1126, 2017.
M. Saha and C. Chakraborty, “Her2Net : A Deep Framework for Semantic Segmentation and Classification of Cell Membranes and Nuclei in Breast Cancer Evaluation,” IEEE Trans. Image Process., vol. 27, no. 5, pp. 2189–2200, 2018.
B. Microbiana et al., “Lung Pattern Classification for Interstitial Lung Diseases Using a Deep Convolutional Neural Network,” IEEE Trans. Med. Imaging, vol. 35, no. 5, pp. 1207–1216, 2016.
Z. Luo, L. Liu, J. Yin, Y. Li, and Z. Wu, “Deep learning of graphs with ngram convolutional neural networks,” IEEE Trans. Knowl. Data Eng., vol. 29, no. 10, pp. 2125–2139, 2017.
A. Krizhevsky, I. Sutskever, and H. Geoffrey E., “ImageNet Classification with Deep Convolutional Neural Networks,” in NIPS, 2012, pp. 1–9.
F. Wu et al., “Temporal interaction and causal influence in community-based question answering,” IEEE Trans. Knowl. Data Eng., vol. 29, no. 10, pp. 2304–2317, 2017.
J. Donahue et al., “Long-Term Recurrent Convolutional Networks for Visual Recognition and Description,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 39, no. 4, pp. 677–691, 2017.
N. Lu, Y. Wu, L. Feng, and J. Song, “Deep Learning for Fall Detection: 3D-CNN Combined with LSTM on Video Kinematic Data,” IEEE J. Biomed. Heal. Informatics, vol. 2194, no. c, 2018. C. Xu, L. Xu, Z. Gao, S. Zhao, H. Zhang, and Y. Zhang, “Direct delineation of myocardial infarction without contrast agents using a joint motion feature learning architecture,” Med. Image Anal., vol. 50, pp. 82–94, 2018.
Jia Deng, Wei Dong, R. Socher, Li-Jia Li, Kai Li, and Li Fei-Fei, “ImageNet: A large-scale hierarchical image database,” in CVPR, 2009, pp. 248–255.
ACKNOWLEDGMENT
The authors would like to thank the anonymous reviewers for their detailed and thoughtful feedback which improved the quality of this paper significantly. This work is fully supported by the National Natural Science Foundation of China Nos. 91846107, 71571058, and 71690235, Anhui Provincial Science and Technology Major Project Nos. 17030801001, and 18030801137, and the Fundamental Research Funds for the Central Universities No. PA2019GDQT0021.
REFERENCES
[1] WHO and ICO, “Human Papillomavirus and Related Diseases Report – WORLD,” HPV Inf. Cent., no. Albania, pp. 1–138, 2014.
[2] R. L. Siegel, K. D. Miller, and A. Jemal, “Cancer statistics, 2018,”
CA. Cancer J. Clin., vol. 68, no. 1, pp. 7–30, 2018.
[3] M. H. Forouzanfar et al., “Breast and cervical cancer in 187
countries between 1980 and 2010: A systematic analysis,” Lancet,
vol. 378, no. 9801, pp. 1461–1484, 2011.
[4] A. I. Ojesina et al., “Landscape of genomic alterations in cervical
carcinomas,” Nature, vol. 506, no. 7488, pp. 371–375, 2014.
[5] C. R. Eheman et al., “National Breast and Cervical Cancer Early
Detection Program data validation project,” Cancer, vol. 120, no.
SUPPL. 16, pp. 2597–2603, 2014.
[6] T. Denkçeken et al., “Elastic light single-scattering spectroscopy for
the detection of cervical precancerous ex vivo,” IEEE Trans.
Biomed. Eng., vol. 60, no. 1, pp. 123–127, 2013.
[7] D. G. Ferris, M. Schiffman, and M. S. Litaker, “Cervicography for
triage of women with mildly abnormal cervical cytology results,”
Am. J. Obstet. Gynecol., vol. 185, no. 4, pp. 939–943, 2001.
[8] H. Greenspan et al., “Automatic detection of anatomical landmarks
in uterine cervix images,” IEEE Trans. Med. Imaging, vol. 28, no. 3,
pp. 454–468, 2009.
[9] A. Milbourne et al., “Results of a pilot study of multispectral digital
colposcopy for the in vivo detection of cervical intraepithelial neoplasia,” Gynecol. Oncol., vol. 99, no. 3 SUPPL., pp. 67–75, 2005.
[10] M. Segondy et al., “Performance of careHPV for detecting high- grade cervical intraepithelial neoplasia among women living with HIV-1 in Burkina Faso and South Africa: HARP study,” Br. J. Cancer, vol. 115, no. 4, pp. 425–430, 2016.
[11] M. F. Mitchell, D. Schottenfeld, G. Tortolero-Luna, S. B. Cantor, and R. Richards-Kortum, “Coloposcopy for the diagnosis of squamous intraepithelial lesions-A meta-analysis,” Obstet. Gynecol., vol. 91, no. 4, pp. 626–631, 1998.
[12] T. Xu et al., “Multi-feature based benchmark for cervical dysplasia classification evaluation,” Pattern Recognit., vol. 63, no. January 2016, pp. 468–475, 2017.
2168-2194 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

Related Posts