Leiden University, Van Steenis Building
Einsteinweg 2, 2333 CC Leiden, room F104
Monday 22 May 2017, 09:00-17:30
The meeting was a great success, in the modern van Steenis Building, with 83 registrations, fine speakers and a lot of networking. See some pictures below:
Below is the background information of the speakers, abstracts and program:
Mining electronic health records for computer-aided decision in ophthalmology
Gwenolé Quellec, Université de Bretagne Occidentale, Inserm, Brest, France
Electronic health records are a major asset for the development of innovative computer-aided decision tools. Through the retrospective analysis of medical records sharing the same clinical interpretations, new decision rules can be mined. Typically, by comparing examination records from patients diagnosed with the same pathology, we can learn to detect that pathology in a new patient, and even to segment its clinical signs in new images. Similarly, by comparing surgeries with similar intraoperative complications, it may be possible to warn the surgeon ahead of time during a new surgery. However, because the most informative data in examination and surgical records are images and videos, the underlying data mining process can be quite challenging. Solutions based on multiple-instance learning and deep learning, two paradigms suitable for image and video mining, will be discussed in this talk. Examples will be given primarily for two ophthalmic applications: retinal pathology screening, using fundus photographs, and cataract surgery monitoring, using surgical microscope videos. Competitive results were achieved by the presented solutions. For instance, our latest algorithm for screening diabetic retinopathy is as good as the top ranking solution of Kaggle’s diabetic retinopathy detection challenge. But, unlike solutions of this challenge, which only provide predictions at the image level, our algorithm can also finely detect lesions in images: it even outperforms lesion detectors trained with manual segmentations. Electronic health record mining is not only convenient, in that it does not require manual segmentations, it is also accurate.
In this talk, I will present two medical imaging problems, histopathology quantification and multi-modal image registration, and discuss the strategies we have developed to come around with the limitations above. First, in the histopathology case we propose an interactive domain adaption method to update random forests predictions with expert feedback. Second, we have recently addressed the problem of multi-modal registration through learning, were our solution relies on data augmentation and the incorporation of learned random forest predictions within a conventional optimization approach.
Dr. Gwenolé Quellec was born in Saint-Renan, France, on November 29, 1982. He received the Ph.D. degree from TELECOM Bretagne, Brest, France, in 2008, and the Habilitation degree from the University of West Brittany, France, in 2015. He was a postdoctoral fellow at the University of Iowa, IA, USA in 2008-2009 and a visiting scholar at the University of Bern, Switzerland, in 2013. He is currently a Research Associate at LaTIM (Medical Information Processing Lab – Inserm UMR 1101), in Brest. His research interests include image and video analysis, as well as machine learning, for medical applications.
Generalizing Convolutions for Deep Learning
Prof. dr. Max Welling, Machine Learning, University of Amsterdam
Arguably, most excitement about deep learning revolves around the performance of convolutional neural networks and their ability to automatically extract useful features from signals. In this talk I will present work from AMLAB where we generalize these convolutions. First we study convolutions on graphs and propose a simple new method to learn embeddings of graphs which are subsequently used for semi-supervised learning and link prediction. We discuss applications to recommender systems and knowledge graphs. Second we propose a new type of convolution on regular grids based on group transformations. This generalizes normal convolutions based on translations to larger groups including the rotation group. Both methods often result in significant improvements relative to the current state of the art.
Joint work with Thomas Kipf, Rianne van den Berg and Taco Cohen.
Prof. Dr. Max Welling is a research chair in Machine Learning at the University of Amsterdam and has a secondary appointments as a senior fellow at the Canadian Institute for Advanced Research (CIFAR). He is co-founder of “Scyfer BV” a university spin-off in deep learning. In the past he held postdoctoral positions at Caltech (’98-’00), UCL (’00-’01) and the U. Toronto (’01-’03). He received his PhD in ’98 under supervision of Nobel laureate Prof. G. ‘t Hooft. Max Welling has served as associate editor in chief of IEEE TPAMI from 2011-2015 (impact factor 4.8). He serves on the board of the NIPS foundation since
2015 (the largest conference in machine learning) and has been program chair and general chair of NIPS in 2013 and 2014 respectively. He was also program chair of AISTATS in 2009 and ECCV in 2016. He has served on the editorial boards of JMLR and JML and was an associate editor for Neurocomputing, JCGS and TPAMI. He received multiple grants from Google, Facebook, Yahoo, NSF, NIH, NWO and ONR-MURI among which an NSF career grant in 2005.
He is recipient of the ECCV Koenderink Prize in 2010. Welling is in the board of the Data Science Research Center in Amsterdam, he directs the Amsterdam Machine Learning Lab (AMLAB), and he co-directs the Qualcomm-UvA deep learning lab (QUVA), the Bosch-UvA Deep Learning lab (DELTA) and the AML4Health Lab.
Max Welling has over 200 scientific publications in machine learning, computer vision, statistics
Congratulations with your PhD, what’s next?
Léon Bemelmans, VinciTech BV
The choice to go for and finalize your PhD reflects a certain feeling with academic subjects, thoroughness end endurance. Continuing the professional career in an academic research environment is not available or desirable for everyone.
The most important alternative seems to be to work at large companies like Philips, ASML, to name only two. However, other options might be to work at smaller companies (Small and Medium Sized Enterprises, SME’s) or even to start your own company.
The talk aims at giving some insights at the consequences for such a choice on aspects like depth of research, variety of tasks and challenges, relation to required skills and remuneration.
These insights are based on the speaker’s experience in founding and running small technically oriented companies (previously Beltech BV, machine vision, 16 f.t.e. and currently VinciTech BV, robotics, 2 f.t.e.). The talk will be followed by a discussion with the audience.
Program of the NVPHBV Spring Meeting 2017
Monday 22nd May 2017, Leiden University, Van Steenis Building, Einsteinweg 2, 2333 CC Leiden, room F104. The meeting is hosted by LIACS and LUMC.
09:30 Opening and welcome
Bart ter Haar Romeny, president NVPHBV
Marius Staring, host LKEB-LUMC
Cor Veenman, host LIACS
09:40 Nearest Neighbor search of points in high dimensional space using binary vantage point features.
Prime Vision B.V.
10:10 Automatic artery/vein classification in retinal images
Fan Huang, Behdad Dashtbozorg, Bart ter Haar Romeny
Medical Image Analysis, Eindhoven University of Technology
11:00 Development of a snapshot hyperspectral imager and automated interpretation algorithms
Nanda van der Stap, Lejla Alic, Klamer Schutte
TNO Intelligent Imaging, Den Haag
11:30 Keynote: The state-of-the-art and future of Intelligent Machines
Machine Learning, University of Amsterdam
12:30 Lunch and Member meeting
14:00 Prediction of histopathological growth patterns in colorectal liver metastases
using a Radiomics approach
Martijn P. A. Starmans 1,2,∗ , Florian E. Buisman 3 , Sebastian R. van der Voort 1,2 , Michel Renckens 2 , Boris Galjart 3 , Pieter M.H. Nierop 3 , Wiro J. Niessen 1,2,4 , Cornelis Verhoef 3 , Jan-Jaap Visser 2 and Stefan Klein 1,2
1 Department of Medical Informatics, Erasmus MC, Rotterdam, NL
2 Department of Radiology and Nuclear Medicine, Erasmus MC, Rotterdam, NL
3 Department of Surgery, Erasmus MC, Rotterdam, NL
4 Faculty of Applied Sciences, Delft University of Technology, NL
14:30 Keynote: Mining electronic health records for computer-aided decision in ophthalmology
Université de Bretagne Occidentale, Inserm, Brest, France
16:00 No Contact Full Body Tremor Measurements
S.L. Pintea, P.J.M.van_Schaik-Bank, J.C. van Gemert, J. Marinus
Computer Vision Lab, Delft University of Technology
16:30 Discussion: Congratulations with your PhD, what’s next?
More information and free registration via http://www.nvphbv.nl
Nearest Neighbor search of points in high dimensional space using binary vantage point features.
Prime Vision B.V. <firstname.lastname@example.org>
A solution of finding nearest neighbors in general metric spaces was needed to retrieve image patches from a large database. Starting point was the vp-tree (vantage point tree) because a 100% retrieval is guaranteed and search speed is fast. Next, the vantage point concept is used to define binary features to indicate if a point is located in a subspace. Binary features can be used for fast testing by using bitwise operators. This is also possible when observation errors occur. A linear search with binary features is fast compared to a normal linear search but slower compared to a vp-tree search. We demonstrate that the binary features can also be used as a basis for a tree search. The binary vantage point tree combines fast tree searching and fast node testing with binary feature testing yielding a speed improvement over normal vp-trees.
Automatic artery/vein classification in retinal images
Fan Huang, Behdad Dashtbozorg, Bart ter Haar Romeny
Eindhoven University of Technology <email@example.com>
Automatic artery/vein (A/V) classification is one of the important topics in retinal image analysis. It allows the researchers to investigate the association between biomarkers and disease progression on a huge amount of data for arteries and veins separately. Recent proposed methods, which employ contextual in- formation of vessels to achieve better A/V classification accuracy, still rely on the performance of pixel-wise classification, which has been received limited attention in recent years. In this paper, we show that these classification methods can be markedly improved. We propose a new normalization technique for extracting four new features which are associated with the lightness reflection of vessels. The accuracy of a Linear Discriminate Analysis (LDA) classifier is used to validate these features. Accuracy rates of 85.1%, 86.9% and 90.6% were obtained on three datasets using only local information. Based on the introduced features the advanced graph- based methods will achieve a better performance on A/V classification.
Development of a snapshot hyperspectral imager and automated interpretation algorithms for laparoscopy
Nanda van der Stap, Lejla Alic, Klamer Schutte
TNO Intelligent Imaging, Den Haag <firstname.lastname@example.org>
Minimization of surgical procedures has led to numerous advantages for the patient. However, limitations to the use of this technique exist as well; identification of critical structures can be challenging for the surgeon due to the lack of tactile feedback and the loss of direct sight. Spectral information may help to improve critical structure identification .
One of the aims of the ECSEL project EXIST is to develop an hyperspectral laparoscope for minimal invasive surgery. Hyperspectral snapshot imaging not only poses challenges for imager design, but also for image formation and -interpretation algorithms. We can report our work-in-progress on developing highly reliable, fast, high-resolution classification algorithms for this hyperspectral laparoscope. The basic classification algorithm is a pixel-wise SVM using ratios of intensities as features, where expert knowledge  has been used to identify sensible ratios. We plan to compare this approach to a Deep Learning-based one, which utilizes spatial information for combined normalization and classification.
This development has been performed in framework of European project EXIST. Schols RM, Laan M ter, Stassen LPS, Bouvy ND, Amelink A, Wieringa FP, et al. Differentiation between Nerve and Adipose Tissue Using Wide-Band (350-1830 nm) in vivo Diffuse Reflectance Spectroscopy. Lasers Surg Med. 2014;46:538–45.
Prediction of histopathological growth patterns in colorectal liver metastases using a Radiomics approach
Martijn P. A. Starmans 1,2,∗ , Florian E. Buisman 3 , Sebastian R. van der Voort 1,2 , Michel Renckens 2 , Boris Galjart 3 , Pieter M.H. Nierop 3 , Wiro J. Niessen 1,2,4 , Cornelis Verhoef 3 , Jan-Jaap Visser 2 and Stefan Klein 1,2 <email@example.com>
1 Department of Medical Informatics, Erasmus MC, Rotterdam, the Netherlands
2 Department of Radiology and Nuclear Medicine, Erasmus MC, Rotterdam, the Netherlands
3 Department of Surgery, Erasmus MC, Rotterdam, the Netherlands
4 Faculty of Applied Sciences, Delft University of Technology, the Netherlands
Anti-angiogenic therapy is crucial in cancer treatment, but the mechanisms limiting efficacy are poorly understood. The Histopathological Growth Pattern (HGP) in ColoRectal Liver Metastases (CRLM) may be a biomarker for (neo)adjuvant treatment response . The HGP is only available after resection. Alternatively, we propose a non-invasive Radiomics approach. We evaluated our method on computed tomography images of 73 patients with two different HGPs. Segmentations were performed by two observers. Support vector machines were used for classification on 182 features reflecting shape, orientation, texture, semantics and first-order statistics. The 95% confidence intervals of the area under the curve were respectively [0.51; 0.74], [0.45: 0.67] and [0.45; 0.68] for the first observer’s segmentations and the second’s two segmentations. Feature selection using an exhaustive grid search increased best case performance to [0.58; 0.81]. Concluding, Radiomics is a promising approach for non-invasive prediction of HGPs in CRLM, but further research is required for clinical use. Sophia Frentzas, Eve Simoneau, Victoria L Bridgeman, Peter B Vermeulen, Shane Foo, Eleftherios Kostaras, Mark R Nathan, Andrew Wotherspoon, Zu-hua Gao, Yu Shi, Gert Van den Eynden, Frances Daley, Clare Peckitt, Xianming Tan, Ayat Salman, Anthoula Lazaris, Patrycja Gazinska, Tracy J Berg, Zak Eltahir, Laila Ritsma, Jacco van Rheenen, Alla Khashper, Gina Brown, Hanna Nystrm, Malin Sund, Steven Van Laere, Evelyne Loyer, Luc Dirix, David Cunningham, Peter Metrakos, and Andrew R Reynolds. Vessel co-option mediates resistance to anti-angiogenic therapy in liver metastases. Nature Medicine, 22(11):1294–1302, October 2016.
No Contact Full Body Tremor Measurements
S.L. Pintea, P.J.M.van_Schaik-Bank, J.C. van Gemert, J. Marinus <firstname.lastname@example.org>
Computer Vision Lab, Delft University of Technology
In the context of hospital tremor analysis, the optimal treatment as well as monitoring patient response to treatment over time, requires objective quantification of relevant motor symptoms. Currently, motor symptoms are evaluated by means of subjective disease-specific clinical examinations and questionnaires, or in some cases employ laboratory-based, time-consuming assessments using multiple body-worn sensors. Such measurements place costly demands on highly qualified personnel and laboratory resources.
We propose: no contact full body tremor measurements. We offer methods for measuring motor functions exclusively from video.
With the astounding progress in Deep Learning, automatic methods now rival human performance in image analysis.
We use deep neural networks in combination with low-level signal processing for measuring tremors from the full body. Our method offers objective tremor measurements and requires no additional sensors; it needs only a video as input. Moreover, it does not require preparation time and provides direct feedback to the doctor and reduces costs considerably.