The Open DDH talks are in the Brakkensteinkamer (route 767) at the Radboud University Medical Centre and start at 14.30.
July 1, 2020
Open DDH: MIDL 2020 edition
Speaker 1: Jasper Linmans
Jasper Linmans is a second-year PhD student at the Radboud University Medical Center in Nijmegen. His research focus lies on uncertainty estimation in the field of digital pathology. In this talk, he will present an efficient approach for out-of-distribution detection in digital pathology using multi-head convolutional neural networks (M-heads).
Abstract: We compare its performance to related and more prevalent approaches, such as deep ensembles, on the task of out-of-distribution (OOD) detection. To this end, we evaluate models trained to discriminate normal lymph node tissue from breast cancer metastases, on lymph nodes containing lymphoma. We show the ability to discriminate between the in-distribution lymph node tissue and lymphoma by evaluating the AUROC based on the uncertainty signal. Here, the best performing multi-head CNN (91.7) outperforms both Monte Carlo dropout (88.3) and deep ensembles (86.8). Furthermore, we show that the meta-loss function of M-heads improves OOD detection in terms of AUROC.
Speaker 2: Maximilian Ilse
Maximilian Ilse is a third-year PhD student at the Amsterdam Machine Learning Lab, led by Max Welling. He is working at the intersection of medical imaging and fundamental machine learning research. His research focus lies on invariance and generalization. In this talk, he will talk about his work on Domain Invariant Variational Autoencoders (DIVA).
Abstract: We consider the problem of domain generalization, namely, how to learn representations given data from a set of domains that generalize to data from a previously unseen domain. We propose the Domain Invariant Variational Autoencoder (DIVA), a generative model that tackles this problem by learning three independent latent subspaces, one for the domain, one for the class, and one for any residual variations. We highlight that due to the generative nature of our model we can also incorporate unlabelled data from known or previously unseen domains. To the best of our knowledge this has not been done before in a domain generalization setting. This property is highly desirable in fields like medical imaging where labeled data is scarce. We experimentally evaluate our model on the rotated MNIST benchmark and a malaria cell images dataset where we show that (i) the learned subspaces are indeed complementary to each other, (ii) we improve upon recent works on this task and (iii) incorporating unlabelled data can boost the performance even further.
In addition, he will talk about what he learned by investigating the limitations of DIVA.
June 3, 2020
Predicting interstitial fibrosis and tubular atrophy in patients with delayed graft function using multiplex immunohistochemistry and deep learning
Abstract: Meyke Hermsen is a PhD student in the Computational Pathology Group and she focuses on deep learning applications for kidney transplant pathology. During the upcoming Open DDH she will discuss her study on early inflammatory infiltrates in delayed graft function patients for the prediction of IFTA development, using multiplex immunohistochemistry and deep learning.
May 6, 2020
Predicting the treatment response to neoadjuvant chemotherapy for breast cancer, a work in progress
Abstract: Witali Aswolinskiy is a postdoctoral researcher in the Computational Pathology Group. He is working in the PROACTING project on predicting the treatment response to neoadjuvant chemotherapy (NACT) for breast cancer patients based on pre-treatment biopsies. The absence of tumor cells in tissue samples after NACT (pathologic complete response (pCR)) is associated with overall and disease-free survival. It allows also for breast-conserving surgery instead of mastectomy. Witali will present his work in progress in the last ten month on the project.
April 1, 2020
Achieving expert-level performance in quantification of retinal pathologies in neovascular age-related macular degeneration using deep learning
Abstract: Bart Liefers is a PhD candidate in DIAG, as part of the A-Eye Research Group. His research is focused on deep learning algorithms for automated retinal image analysis as part of the AMI-Ophthalmology project. In particular, it is focused on quantification of retinal abnormalities in age-related macular degeneration (AMD). AMD is a leading cause of blindness for elderly people, that can be treated in some cases with intra-ocular injections. These injections are costly and need to be administered frequently, and therefore form a significant burden on healthcare and patient. Furthermore, close monitoring of disease activity is required, which is performed by three-dimensional imaging of the retina, using optical coherence tomography (OCT). Manual interpretation of these images is subjective and time consuming. In his latest work, Bart presents a deep learning model that is able to provide volumetric quantification of thirteen retinal abnormalities. The proposed model accelerates interpretation of OCT volumes and surpasses manual reading, both in terms of attainable level of extracted information and consistency. This can potentially lead to a reduction of costs in interpretation of clinical trials and improve personalized clinical care.
March 4, 2020
Extending unsupervised neural image compression with supervised multitask learning.
Abstract: David is a 4th year PhD student in the Computational Pathology Group. His research is mainly focused on building deep learning models that can analyze entire histopathological images at once, i.e., predicting patient-level labels from gigapixel images. Last summer, he published his work on "unsupervised neural image compression" in TPAMI, a method that reduces the size of gigapixel images while retaining high-level features. Today, he is going to explain this method and introduce the latest extension of his work "Extending unsupervised neural image compression with supervised multitask learning", currently under review for MIDL2020.
February 5, 2020
Deep learning in CT Thorax Abdomen: Organ detection and segmentation.
Gabriel Efrain Humpire Mamani
Abstract: In this presentation, Gabriel will summarize the projects he worked on his PhD at DIAG as part of the AMI Oncology project. He worked on organ detection, spleen segmentation, kidney segmentation, and multi-organ segmentation on thorax-abdomen CT scans. He will explain the current experiments he is executing using transfer learning.
December 4, 2019
Ultrasound imaging in resource-limited settings by combining a low-cost ultrasound device with artificial intelligence.
Thomas van den Heuvel
Abstract: Worldwide, 99% of all maternal deaths occur in resource-limited countries. Ultrasound imaging is a widely used technique to detect maternal and fetal risk factors and recently, ultrasound devices have become cheaper and more portable. Unfortunately, ultrasound imaging often remains out of reach for pregnant women in developing countries, because it requires a trained sonographer to acquire and interpret the ultrasound images, while there is a severe shortage of sonographers in these countries. In this presentation, Thomas will explain how he wants to introduce ultrasound imaging in resource-limited settings by combining a low-cost ultrasound device with artificial intelligence.
November 7, 2019
Automated assessment of lymph node status from colorectal and breast cancer resections using deep neural networks
Abstract: In this presentation, Péter will briefly summarize all the projects he has covered during his PhD at DIAG, including the CAMELYON17 challenge, the CAMELYON dataset, his work within the AMI project, whole-slide image registration, the digital pathology library, and resolution-agnostic whole-slide image segmentation. Additionally, he will explain in detail his last project, where he explores the possibilities of transfer learning in the field of cancer metastasis detection in lymph nodes.
June 5, 2019
Dense Segmentation in Selected Dimensions: Application to Retinal Optical Coherence Tomography
Abstract: During this week's Open DDH shall Bart present a novel convolutional neural network architecture designed for dense segmentation in a subset of the dimensions of the input data. The architecture takes an N-dimensional image as input, and produces a label for every pixel in M output dimensions, where 0 < M < N. Large context is incorporated by an encoder-decoder structure, while funneling shortcut subnetworks provide precise localization. Bart and colleagues demonstrate applicability of the architecture on two problems in retinal optical coherence tomography: segmentation of geographic atrophy and segmentation of retinal layers. Performance is compared against two baseline methods, that leave out either the encoder-decoder structure or the shortcut subnetworks. For segmentation of geographic atrophy, an average Dice score of 0.49 ± 0.21 was obtained, compared to 0.46 ± 0.22 and 0.28 ± 0.19 for the baseline methods, respectively. For the layer-segmentation task, the proposed architecture achieved a mean absolute error of 1.305 ± 0.547 pixels compared to 1.967 ± 0.841 and 2.166 ± 0.886± for the baseline methods. This work is included in the Proceedings of Machine Learning Research and shall be presented at MIDL 2019.
May 8, 2019
New treatment paradigms for improved cancer treatment with radiotherapy
Abstract: Cancer is the second leading cause of death worldwide. Radiation therapy, or radiotherapy (RT) for short, plays a pivotal role in the treatment of many cancers, where approximately 50% of cancer patients can benefit from RT in the management of their disease. During radiotherapy ionizing radiation, generally produced by a linear accelerator (linac) is delivered with the intent of killing malignant cells. To limit side effects, and to exploit the higher repair capacity of normal tissue to compared to tumor cells, the total radiation dose is typically delivered in smaller daily portions over a period of several weeks.
Historically radiotherapy has been one of the most technologically advanced sub fields of medicine with a strong interplay between clinicians and physicists where there is a lot of unlocked potential for the application of deep learning. In this talk I will give an overview of what radiotherapy is, what a typical workflow looks like and how deep learning can help in improving the care for cancer patients. In particular I will talk about the use of segmentation for treatment planning, deep learning for image improvement and reconstruction and what I will be working on the next couple of years by using deep learning to enable new treatment paradigms to improve treatment with radiotherapy.
April 3, 2019
Learning-based vertebra segmentation, identification and partitioning
Abstract: The spine is visualized on many CT and MR exams, including thorax and abdomen scans that were originally not intended for spine imaging. Because these often cover several but not all vertebrae, it is difficult to make strong assumptions for automatic analysis. Challenges are therefore the unknown number of target structures (vertebrae) in the image, their anatomical identification (which vertebrae are visible? must not assign the same label to two vertebrae) and that some biomarkers are related only to part of the vertebrae, often the vertebral body. This talk covers an instance segmentation approach for vertebra detection, segmentation, and anatomical identification, and a partitioning approach to separate vertebral body and arch based on thin-plate spline surfaces positioned by a convolutional neural network.
March 27, 2019
From Researcher to “Defective and Diseased Alien”
Bram Platel received his doctorate from the Eindhoven University of Technology in 2007. He started to work at DIAG in 2010 with Nico Karssemeijer, mostly working on CAD for mammography; later-on Bram’s focus shifted towards image analysis for neurological disorders. Following strange symptoms at the end of 2013, he was diagnosed with Multiple Sclerosis (MS). After two years, Bram was forced to reduce his working hours to a minimum. In February of 2016, on a routine MRI-scan for his MS, multiple brain tumors were found. To treat this unique form of lymphoma, Bram received various chemotherapy treatments during a period of half a year. The treatment ended with a blood stem cell transplantation. Coincidentally, abroad a similar therapy is recently being used to treat aggressive forms of multiple sclerosis. Since the treatment, Bram's lymphoma is in complete remission, and his MS hasn't progressed. In this non-scientific presentation, Bram will talk about this period, his new foundation MS in beeld, and his emigration to the Philippines next month.
March 6, 2019
Inverse problems in medical imaging
Abstract: Inverse problem is the type of problems in natural sciences when one has to infer from a set of observations the causal factors that produced them. In medical imaging, important examples of inverse problems would be reconstruction in CT and MRI, where the volumetric representation of an object is computed from the projection and Fourier space data respectively. In a classical approach, one relies on domain specific knowledge contained in physical-analytical models to develop a reconstruction algorithm, which is often given by a certain iterative refinement procedure. Recent research in inverse problems seeks to develop a mathematically coherent foundation for combining data driven models, based on deep learning, with the analytical knowledge contained in the classical reconstruction procedures. In this talk we will give a brief overview of these developments and then focus on particular applications in Digital Breast Tomosynthesis and MRI reconstruction.