10:40
MS53: Advanced Numerical Methods in Image Processing (Part 2)
Chair: Fritz Keinert
10:40
25 mins
|
A Learning-Based Formulation of Parametric Curve Fitting for Bioimage Analysis
Soham Mandal, Virginie Uhlmann
Abstract: Parametric curve models are convenient to describe and quantitatively characterize the contour of objects in bioimages. Unfortunately, designing algorithms to fit smoothly such models onto image data classically requires significant domain expertise. Here, we propose a convolutional neural network-based approach to predict a continuous parametric representation of the outline of biological objects. We successfully apply our method on the Kaggle 2018 Data Science Bowl dataset composed of a varied collection of images of cell nuclei. This work is a first step towards user-friendly bioimage analysis tools that extract continuously-defined representations of objects.
|
11:05
25 mins
|
Deformable models and convolutional neural networks for cardiac image segmentation…. which approach for a successful application in the clinical practice?
Claudio Fabbri, Davide Borra, Cristiana Corsi
Abstract: In the scenario of precision medicine, the availability of tools to support clinical decision making is crucial. In the area of cardiac diseases, diagnosis is based on clinical indexes able to provide information about cardiac morphology and function. These parameters are mainly derived from image processing. Several studies have proposed automated and semi-automated methods for cardiac chamber segmentation to extract endocardial and epicardial surfaces. Some common strategies are related to morphology approximations for the cardiac structures [1], but the need for a priori knowledge of the shape of the heart limits the usefulness of this approach in patients with irregular/unusual cardiac anatomy. Other methods are based on statistical distribution of the gray level intensity, which allows the characterization of different regions in the image, or on edge-based techniques, which drive the evolution of a level set function until the myocardial borders are detected. The combination of these two approaches has also shown encouraging results [2][3]. Recently, advances in machine learning techniques based on convolutional neural networks have attracted the attention of the research community in the biomedical image processing field and several promising applications have been proposed in cardiac imaging [4].
In this work we review these approaches applied to data from magnetic resonance imaging and present two models for left heart segmentation: a region-based segmentation using statistical properties of the image and an edge-based approach [5], and convolutional neural networks [6]. These approaches are applied to data from patients affected by different pathologies including heart failure and atrial fibrillation. Clinical indexes are then extracted and validated against parameters derived from manual annotations of data by experts.
References
[1] M Lorenzo-Valdés, GI Sanchez-Ortiz, AG Elkington, RH Mohiaddin, D Rueckert, Segmentation of 4D cardiac MR images using a probabilistic atlas and the EM algorithm Medical Image Analysis 2004, 8(3):255–65.
[2] C Li, X Jia. Y Sun, Improved semi-automated segmentation of cardiac CT and MR images International Symposium on Biomedical Imaging 2009, 2:25–28.
[3] C Corsi, F Veronesi, C Lamberti, D Bardo, EB Jamison, RM Lang, V Mor-Avi, Automated frame-by-frame endocardial border detection from cardiac magnetic resonance images for quantitative assessment of left ventricular function: validation and clinical feasibility Journal of Magnetic Resonance Imaging 2009;29(3):560–8.
[4] D Dey, PJ Slomka, P Leeson, D Comaniciu, S Shrestha, PP Sengupta, TH Marwick, Artificial Intelligence in Cardiovascular Imaging: JACC State-of-the-Art Review, J Am Coll Cardiol. 2019;73(11):1317-133.
[5] C Fabbri, K Kawaji, N Nazir, V Mor-Avi, A Patel, C Corsi, A Semiautomated Approach for the Quantification of the Left Ventricle Chamber Volumes From Cine Magnetic Resonance Images, in Computing in Cardiology 2018 (Maastricht, The Netherlands, 23-26 September 2018), IEEE Press 45:1-4.
[6] D Borra, A Masci, L Esposito, A Andalò, CFabbri, C Corsi, A Semantic-Wise Convolutional Neural Network Approach for 3-D Left Atrium Segmentation from Late Gadolinium Enhanced Magnetic Resonance Imaging, in: Mihaela Pop, Maxime Sermesant, Jichao Zhao, Shuo Li, Kristin McLeod, Alistair Young, Kawal Rhode, Tommaso Mansi (Eds.), Statistical Atlases and Computational Models of the Heart. Atrial Segmentation and LV Quantification Challenges, Lecture Notes in Computer Science 2019;11395:329-338.
|
11:30
25 mins
|
Deep learning techniques for fluorescence imaging: detection and brightness estimation
Andrea Samorè
Abstract: Deep learning techniques have recently revolutionized the imaging field, often achieving impressive results and human level performance on complex computer vision tasks.
After a brief introduction and an overview on the state of the art, we will consider applications of deep learning techniques to biomedical imaging and the associated peculiar difficulties, with a case study on fluorescence image analysis.
For this problem, object detection is generally required to assess the presence of a molecule of interest, while brightness estimation is often performed to measure its concentration [1].
A new neural network for object detection and brightness estimation, which allows for completely automated fluorescence image analysis without requiring any input parameter from the user during the inference phase, will be presented. Additionally, methods to assemble a good dataset with a limited number of images, as is typical with clinical data, will be discussed, together with strategies to reduce the computational load of the network through quantization and model selection for deployment on resource constrained embedded systems.
References:
[1] Lazzaro, Damiana; Morigi, Serena; Melpignano, P.; Loli Piccolomini, E.; Benini, Luca, Image enhancement variational methods for Enabling Strong Cost Reduction in OLED-based Point- of-Care Immunofluorescent Diagnostic Systems, «INTERNATIONAL JOURNAL FOR NUMERICAL METHODS IN BIOMEDICAL ENGINEERING», 2018, 34, pp. 2932 - 2951
|
11:55
25 mins
|
Semi-automatic quantification of invasion and migration from images of cancer cells
Marilisa Cortesi, Caroline Ford, Emanuele Giordano
Abstract: The quantification of invasion and migration, fundamental aspects of cancer cell behaviour used to evaluate the effectiveness of potential new treatments, often relies on manual cell counting. This considerably limits the accuracy of these assays and the number of samples that can be considered. As microscopy images are often acquired, to allow for later or repeated analyses, we developed a software tool for the automatic recognition of cancer cells in images from transwell assays [1].
This tool relies on the marker controlled watershed transform to segment the images and identify cells and integrates two different approaches (morphological filtering and machine learning using support vector machines) to isolate and count the objects of interest. Our method was validated on a dataset comprising 180 images from two different ovarian cancer cell lines characterized by different morphologies, that was completely independent from the training set (composed of 160 images of the same cell lines). The results obtained with our software were compared with manual counting and the dependence on the operator was evaluated considering six users with different levels of expertise with this task.
The use of our software was shown to improve the robustness and reproducibility of the results, reducing their dependence on the operator and their individual expertise, as well as facilitating the analysis of high density images.
References
[1] Cortesi M, Llamosas E, Henry C, Kumaran R, Ng B, Youkhana J, Ford C. I-AbACUS: a reliable software tool for the semi-automatic analysis of invasion and migration transwell assays. Scientific Reports (2018) DOI:10.1038/s41598-018-22091-5
|
|
