Baidu Research

<!doctype html> <html> <head> <meta charset="utf-8"> <meta name="renderer" content="webkit" /> <meta name="format-detection" content="telephone=no"/> <meta http-equiv="X-UA-Compatible" content="IE=edge,chrome=1" /> <meta content="width=device-width, initial-scale=1, maximum-scale=1, minimum-scale=1" name="viewport" /> <title>Baidu Research</title> <link rel="shortcut icon" href="../web/images/favicon.png" type="image/x-icon"> <link rel="stylesheet" href="../web/css/reset.css"> <link rel="stylesheet" href="../web/css/animate.css"> <link rel="stylesheet" href="../web/css/Bootstrap.css"> <link rel="stylesheet" href="../web/css/style.css"> <link rel="stylesheet" href="../web/css/media.css?3333"> <script src="../web/js/jquery.min.js"></script> </head> <body class="n_body" ondragstart="return false">  <div class="n_header"> <div class="container"> <div class="header01"> <div class="logo"> <a class="h_logo" href="/Index" style="background-image: url(../web/images/logo.png);"><img src="../web/images/logo.png" alt="Baidu Research"></a> </div> <div class="nav"> <ul> <li><a href="/Index">Home</a></li> <li><a href="/Publications">Publications</a></li> <li><a href="/Research_Areas?id=55">Research Areas</a> <div class="nav_er"> <ul class="div_dl "> <li><a href="/Research_Areas/index-view?id=55">Data Science and Data Mining</a></li> <li><a href="/Research_Areas/index-view?id=56">Natural Language and Speech</a></li> <li><a href="/Research_Areas/index-view?id=57">Business Intelligence</a></li> <li><a href="/Research_Areas/index-view?id=58">Robotics and Autonomous Driving</a></li> <li><a href="/Research_Areas/index-view?id=59">Computer Vision</a></li> <li><a href="/Research_Areas/index-view?id=60">Machine Learning and Deep Learning</a></li> <li><a href="/Research_Areas/index-view?id=61">Computational Biology and Bioinformatics</a></li> <li><a href="/Research_Areas/index-view?id=62">High Performance Computing</a></li> <li><a href="/Research_Areas/index-view?id=75">Quantum Computing</a></li> </ul> </div> </li> <li><a class="active" href="/Blog">Blog</a></li> <li><a href="/Career">Careers</a></li> <li><a href="/Downloads">Downloads</a></li> <li><a href="/AI_Colloquium">AI Colloquium</a></li> </ul> <div class="nav_btn"><span></span></div> </div> </div> </div> <div class="header03 "> <div class="logo"><a href="/Index"><img src="../web/images/logo.png" alt="Baidu Research"></a></div> <div class="nav"> <ul> <li><a href="/Index">Home</a></li> <li><a href="/Publications">Publications</a></li> <li><a href="/Research_Areas?id=55">Research Areas</a> <div class="nav_er"> <ul class="div_dl "> <li><a href="/Research_Areas/index-view?id=55">Data Science and Data Mining</a></li> <li><a href="/Research_Areas/index-view?id=56">Natural Language and Speech</a></li> <li><a href="/Research_Areas/index-view?id=57">Business Intelligence</a></li> <li><a href="/Research_Areas/index-view?id=58">Robotics and Autonomous Driving</a></li> <li><a href="/Research_Areas/index-view?id=59">Computer Vision</a></li> <li><a href="/Research_Areas/index-view?id=60">Machine Learning and Deep Learning</a></li> <li><a href="/Research_Areas/index-view?id=61">Computational Biology and Bioinformatics</a></li> <li><a href="/Research_Areas/index-view?id=62">High Performance Computing</a></li> <li><a href="/Research_Areas/index-view?id=75">Quantum Computing</a></li> </ul> </div> </li> <li><a href="/Blog">Blog</a></li> <li><a href="/Career">Careers</a></li> <li><a href="/Downloads">Downloads</a></li> <li><a href="/AI_Colloquium">AI Colloquium</a></li> </ul> </div> <div class="nav_btn"><span></span></div> </div> </div> <div class="baidu-page-banner blog-side" style="background: url(/Public/uploads/5ae96c0a7676c.png);"> <div class="container"> <div class="baidu-page-title wow fadeIn">Blog</div> </div> </div> <div class="content-info"> <div class="container-details-er"> <div class="blog-details-title">Measuring Lesion Geometry with AI</div> <div class="blog-details-date"><p>2019-03-01</p><a href="/Blog">Back to list</a></div> <p style=";font-size: medium;font-family: Calibri, sans-serif;white-space: normal;text-align: justify"><span style="font-size: 19px">Lesions characterized by computed tomography (CT) scans, are arguably often elliptical objects. Therefore, radiologists often use two perpendicular diameters to measure the geometry of lesions<sup>1</sup>, which closely represent the major and minor axises of an ellipse. Figure 1 shows an example of a lesion in a CT scan and its corresponding two-axis annotation.</span></p><p style=";font-size: medium;font-family: Calibri, sans-serif;white-space: normal;text-align: center"><img src="/ueditor/upload/20190301/1551428345853424.jpg" title="1551428345853424.jpg" alt="1.jpg"/></p><p style=";font-size: medium;font-family: Calibri, sans-serif;white-space: normal;text-align: center"><span style="font-size: 19px">Figure 1.</span></p><p style=";font-size: medium;font-family: Calibri, sans-serif;white-space: normal;text-align: justify"><span style="font-size: 19px"> </span></p><p style=";font-size: medium;font-family: Calibri, sans-serif;white-space: normal;text-align: justify"><span style="font-size: 19px">To aid radiologists in efficiently reviewing these CT scans, lots of object detection algorithms originally developed for analyzing nature images, e.g. Region Proposal Network (RPN), have been adopted to detect lesions in medical images. However, most of these algorithms only predict the bounding box of a lesion without fully measuring the elliptical geometry of the lesion, since objects in nature images are mostly annotated with bounding boxes. Extensions on RPN have been introduced to also predict the rotation angle of the object, but it has its own limitations. Figure 2 shows two illustrative examples.</span></p><p style=";font-size: medium;font-family: Calibri, sans-serif;white-space: normal;text-align: center"><span style="font-size: 19px"><img src="/ueditor/upload/20190301/1551428355402406.png" title="1551428355402406.png" alt="2.png" width="435" height="212"/></span></p><p style=";font-size: medium;font-family: Calibri, sans-serif;white-space: normal;text-align: justify"><span style="font-size: 19px">Figure 2. Two illustrative examples of how the rotation angle may affect the overlap between two ellipses differently. The ellipse of solid line is the ground truth and the ellipse of dash line is proposed by the detection model, where both ellipses only differ in the rotation angle. (a) the major axis is significantly longer than the minor axis, (b) the major axis is slightly longer than the minor axis.</span></p><p style=";font-size: medium;font-family: Calibri, sans-serif;white-space: normal;text-align: justify"><span style="font-size: 19px"> </span></p><p style=";font-size: medium;font-family: Calibri, sans-serif;white-space: normal;text-align: justify"><span style="font-size: 19px">We present Gaussian Proposal Network (GPN) that learns to measure the elliptical geometry of lesions by modelling their bounding ellipses as 2D Gaussian distributions on the image plane. Figure 3 shows an illustrative comparison between GPN and RPN. Instead of predicting the rotation angle, GPN minimizes the Kullback-Leibler (KL) divergence between the proposed Gaussian distribution and the ground truth Gaussian distribution to jointly localize the lesion and predict its elliptical geometry. </span></p><p style=";font-size: medium;font-family: Calibri, sans-serif;white-space: normal;text-align: center"><img src="/ueditor/upload/20190301/1551428369136986.png" title="1551428369136986.png" alt="3.png" width="484" height="396"/></p><p style=";font-size: medium;font-family: Calibri, sans-serif;white-space: normal;text-align: center"><span style="font-size: 19px">Figure 3.</span></p><p style=";font-size: medium;font-family: Calibri, sans-serif;white-space: normal;text-align: justify"><span style="font-size: 19px">We tested our algorithm on the large scale medical imaging dataset DeepLesion recently released from NIH, which has around 32,000 CT scans of lesions with various sizes and shapes. GPN proposes lesion bounding ellipses of much higher overlap with the ground truth than RPN. In terms of lesion localization score (FROC), GPN outperforms RPN by more than 10%. Figure 4 shows a few examples. </span></p><p style=";font-size: medium;font-family: Calibri, sans-serif;white-space: normal;text-align: center"><img src="/ueditor/upload/20190301/1551428382808036.png" title="1551428382808036.png" alt="4.png" width="482" height="222"/></p><p style=";font-size: medium;font-family: Calibri, sans-serif;white-space: normal;text-align: justify"><span style="font-size: 19px">Figure 4. Selected examples of top 3 proposed bounding ellipses from GPN (cyan) and RPN (magenta) compared to the ground truth (orange) on the test set of DeepLesion.</span></p><p style=";font-size: medium;font-family: Calibri, sans-serif;white-space: normal;text-align: justify"><span style="font-size: 19px">For more details, please refer to our paper (https://arxiv.org/abs/1902.09658). We also open sourced our algorithm at github</span></p><p style=";font-size: medium;font-family: Calibri, sans-serif;white-space: normal;text-align: justify"><span style="font-size: 19px">(https://github.com/baidu-research/GPN) for reproducing our results and further research on this topic.</span></p><p style=";font-size: medium;font-family: Calibri, sans-serif;white-space: normal;text-align: justify"><span style="font-size: 19px"> </span></p><p style=";font-size: medium;font-family: Calibri, sans-serif;white-space: normal;text-align: justify"><span style="font-size: 19px">Authors: Yi Li</span></p><p><span style="font-size: 19px;font-family: Calibri, sans-serif">The technical term of this annotation is called response evaluation criteria in solid tumors (RECIST) diameters. Each RECIST diameter annotation consists of two axies: the first one measures the longest diameter of the lesion and the second one measures the longest diameter perpendicular to the first axis.</span></p><p><br/></p> <div class="pager"> <a href="/Blog/index-view?id=112"> <i class="glyphicon glyphicon-menu-up"></i>Previous One：Baidu Team Wins the Suggestion Mining Challenge at SemEval 2019 </a> <a href="/Blog/index-view?id=110"> <i class="glyphicon glyphicon-menu-down"></i>Next One：Baidu Teams up with Academia, awarding Grand Prizes for Breakthroughs in the ICDAR 2019 Competitions </a> </div> </div> </div> <footer> <div class="baidu-bottom"> <div class="container"> <div class="col-md-6 col-xs-12"> <h2>Baidu Research</h2> <p>1195 Bordeaux Drive Sunnyvale, CA 94089<br>Baidu Technology Park, No. 10 Xibeiwang East Road, Haidian District, Beijing, China<br>Media Inquiries: <a href="mailto:intlcomm@baidu.com">intlcomm@baidu.com</a><br>General Inquries: <a href="mailto:air-info@baidu.com">air-info@baidu.com</a></p> </div> <div class="col-md-6 col-xs-12"> <ul class="social-icons"> <li><a href="https://twitter.com/baiduresearch" target="_blank"><img src="../web/images/ico-2.png"></a> </li> <li><a href="https://www.linkedin.com/company/baidu-usa" target="_blank"><img src="../web/images/ico-3.png"></a> </li> </ul> <div class="baidu-weibu"> <div class="baidu-img"><img src="../web/images/f-logo.png"></div> <div class="baidu-links"> <a class="baidu-links-title" href="javascript:;">Links</a> <ul class="baidu-links-friends"> <li><a href="http://ai.baidu.com/" target="_blank">Baidu AI Open Platform</a> </li> <li><a href="http://www.dlnel.org/" target="_blank">DLNEL</a> </li> </ul> </div> </div> </div> </div> </div> <div class="baidu-foot">© 2018 Baidu Research</div> </footer> <script src="../web/js/bootstrap.min.js"></script> <script src="../web/js/wow.js"></script> <script src="../web/js/base.js"></script> </body> </html>

CINXE.COM

Baidu Research