CBORT 2021 Symposium

 

On May 7th, 2021, the Center for Biomedical OCT Research and Translation (CBORT) hosted a one day symposium on Machine Learning in OCT Imaging: Novel Applications & New Research Opportunities.

Machine Learning is a powerful tool and has shown great promise for enhanced visualization in numerous medical imaging modalities including Computed Tomography, Magnetic Resonance Imaging, and Ultrasound. Machine Learning in OCT Imaging aimed to highlight the increasing impact of advanced machine learning on biomedical optics and in particular Optical Coherence Tomography (OCT). In addition, the program served to connect the biomedical optics community and machine learning / computer vision experts.

Recordings of the symposium’s sessions with select contributions are available below.

Session 1: Machine Learning in Medical Imaging

State of the Art & Open Problems in Machine Learning for Imaging

Daniel Rueckert, Imperial College London

We will give an overview of the current state-of-the-art in machine learning for medical imaging applications such as reconstruction, segmentation and classification. In particular, we will illustrate deep learning approaches based on Convolutional Neural Networks (CNN). We will focus on deep learning models that use encoder-decoder networks and show these can be used for image analysis tasks. We show some applications of CNNs in the context of image classification. Finally, we will discuss some open challenges for deep learning approaches such as explainability and verification of deep learning.

https://www.youtube.com/watch?v=PKJAQFnyCW

Lessons from Machine Learning for MR Data Acquisition, Image Reconstruction & Analysis

Florian Knoll, New York University

Recent basic science developments in optimization and machine learning, as well as widespread access to powerful computing resources and large datasets have the potential to substantially change the way magnetic resonance imaging is performed. I will discuss the potential of these developments to make imaging faster, cheaper, easier to use, more patient friendly and accessible, and to obtain new information. I will cover both methodological developments as well as clinical translation and validation, and will discuss ongoing developments as well as currently open research questions and potential pitfalls of the methodology.

https://www.youtube.com/watch?v=PKJAQFnyCW

Growth of Machine Learning & Its Applications in OCT

Nishant Mohan, PhotoniCare Inc.

The last decade has seen an unprecedented growth in capabilities of artificial intelligence (AI) and machine learning (ML) based techniques. In this presentation we will discuss driving factors behind this growth and explore key machine learning paradigms. We will discuss how OCT uniquely benefits from ML and key application categories related to successful application of ML in OCT. Practical issues of employing ML techniques in medical devices from commercial and regulatory perspectives will also be explored.

https://www.youtube.com/watch?v=PKJAQFnyCW

Session 2: The Potential for Machine Learning in OCT

Quirks and Twists of OCT Imaging for the Computational Scientist

Nestor Uribe-Patarroyo

In this presentation, we will discuss the basic principles of OCT and how they determine the many intrinsic and informative  properties of OCT imaging. We will revisit key concepts of coherent imaging, including speckle, light scattering and attenuation, and how each impacts image formation and interpretation. Existing opportunities to improve structural OCT imaging with deep learning will be explored, with special consideration for the physics and constraints  unique to this modality. Finally, we will provide  a brief introduction to OCT functional imaging, along with current challenges which stand to benefit from deep learning approaches.

https://www.youtube.com/watch?v=nZvqbmYdIjE

Developing machine learning algorithms for disease detection with clinical impact and utility

Lida Hariri & Markus Herrmann

Machine learning is a powerful tool for enhanced visualization and assessment of macroscopic imaging modalities, such as computed tomography, and microscopic imaging modalities, such as digitized histology slides and optical coherence tomography, for disease detection and monitoring. Significant consideration must be given to the study design and approach to ensure there is appropriate clinical relevance, utility, and accuracy for each specific clinical application. In this talk, we will discuss these topics in detail, using our work developing machine learning algorithms for detection of fibrotic pulmonary disease as an example.

No recording available.

Computer-Aided Detection of Suspicious Tissue in Wide-Field OCT Image Data of the Breast Using Convolutional Neural Networks

Vladimir Pekar, Perimeter Medical

We present an algorithm for computer-aided detection of suspicious areas in WF-OCT scans of surgical margins in breast-conserving surgery (BCS), using Convolutional Neural Networks (CNNs). A simple CNN architecture has been designed to classify square overlapping image patches extracted from WF-OCT B-scans as suspicious or non-suspicious. Classification only takes a few seconds per margin consisting of several hundred B-scans on a modern GPU-powered workstation. Suspicious patches from adjacent B-scans are grouped as clusters which are then reviewed using a specialized graphical user interface (GUI). The developed method allowed to both reduce speed and improve the accuracy of margin assessment when evaluated by trained non-clinical users in a pilot internal study. Standalone classification performance of the proposed algorithm on a representative test set is presented as well as results of a reader study, where several clinical users performed algorithm-assisted data review in a clinically relevant timeframe. The results of the performance evaluation study suggest that the developed algorithm is a promising tool to be used intraoperatively for efficient review of WF-OCT data of surgical margins in BCS.

https://www.youtube.com/watch?v=nZvqbmYdIjE

Session 3: Machine Learning in Biomedical Optics

Machine Learning Methods for Automated Detection of Regions of Fibrosis and Adipose within OCT Images of the Heart

Christine Hendon, Columbia University

Cardiac arrhythmias are a major source of morbidity and mortality in the United States, where it is estimated that millions of people have arrhythmias that cannot be controlled with medications or devices. Tissue composition within the heart plays a critical role in the pathology of cardiovascular disease, tissue remodeling, and arrhythmogenic substrates. My group, the Structure Function Imaging Laboratory at Columbia University, aims to develop optical coherence tomography technologies for real-time assessment of interventional procedures for generation of substrate maps for procedural guidance. Within this talk I will review a range of methods that we have developed to segment and classify areas of adipose, fibrotic myocardium, collagen tissue and normal myocardium within human donor hearts. These range from machine learning methods such as relevance vector machines to artificial intelligence methods such as convolutional neural networks. I will also highlight our approaches to challenges in developing supervised learning methods due to imperfections in the manual labeling process.

https://www.youtube.com/watch?v=UWTBMwdIbBg

Deep Learning for Automated Segmentation of Coronary Anatomical Layers and Atherosclerotic Lesions with Intravascular Polarimetry

Mohammad Haft-Javaherian, Massachusetts General Hospital

Coronary artery disease is one of the leading causes of death, and nearly one-third of the patients who experience a coronary episode will die in the same year. The narrowing or occlusion of the coronary artery compromises the oxygen and nutrient delivery to the heart. While angiography is the standard imaging modality for stenosis quantification, polarization-sensitive optical frequency domain imaging (PS-OCT) provides high-resolution images of subsurface microstructures of OCT in addition to the quantitative metrics related to tissue composition. However, its wide clinical adoption has been impeded by the subjective criteria applied traditionally to OCT image interpretation by the larger interventional cardiology community. These qualitative classifications limit the uptake and utility of OCT. This study proposed a convolutional neural network model and optimized it using a new multi-term loss function to classify the lumen, intima, media, and plaques. Our multi-class classification model outperforms the state-of-the-art methods in detecting the anatomical layers based on accuracy, Dice coefficient, and average boundary error.

https://www.youtube.com/watch?v=UWTBMwdIbBg

GANs and Synthetic Data for Deep Learning Colonoscopy Image Analysis

Nick Durr, Johns Hopkins University

Like many other medical imaging problems, colonoscopy has enormous potential to benefit from automated image analysis with deep learning but also presents formidable challenges in acquiring suitable ground-truth data for training. This presentation discusses some strategies to overcome these challenges by: (1) creating synthetic datasets, (2) implementing domain adaptation tools, and (3) exploiting unlabeled data. Lastly, I will describe some recent efforts in hardware and algorithm co-design to create optical systems that are optimized for deep learning analysis.

https://www.youtube.com/watch?v=UWTBMwdIbBg

Adaptive Deep Learning for Imaging in Scattering Media

Lei Tian, Boston University

Deep learning has been broadly applied to imaging in scattering applications. A common limitation is that the performance of the neural network heavily relies on matching the scattering conditions during the training and testing times. In this talk, I will discuss a dynamic synthesis approach for adaptive descattering and demonstrate superior imaging performance across a wide range of scattering conditions.

https://www.youtube.com/watch?v=UWTBMwdIbBg

Session 4: Machine Learning in Ophthalmic OCT

Machine Learning for Optimal Utilization of OCT in Laboratory and Clinical Trials

Sina Farsiu, Duke University

While OCT image segmentation algorithms’ performance has been significantly improved in recent years, a critical question has not yet been addressed: “How far is segmentation performance from its theoretical limit?” The answer to this question justifies further investments of time and financial resources to gain higher segmentation accuracy. In this talk, we review optical, mathematical, and biological factors that limit accurate quantification of functional and anatomical features in biomedical OCT Imaging applications. We introduce machine learning and computational imaging software and hardware to enhance the resolution and accuracy of quantifying biomarkers of disease in various laboratory and clinical applications of OCT and adaptive optics OCT systems.

No recording available.

Applying Unsupervised Machine Learning to Assess OCT Scans for Glaucoma

Mengyu Wang, Massachusetts Eye and Ear Infirmary

Glaucoma is an optic neuropathy resulting in irreversible vision loss. In this work, we determined representative structural patterns of the retinal structure with unsupervised machine learning from OCT scans. The structural patterns of the retinal structure were subsequently associated with glaucoma diagnostic parameters. Our results suggest that the structural patterns of the retinal structure determined by unsupervised machine learning can be used to improve our understanding of glaucomatous damages in retinal structure as well glaucoma diagnostic accuracy.

https://www.youtube.com/watch?v=_VOueRtahFI

Artificial Intelligence in OCT Angiography

Yali Jia, Oregon Health & Sciences University

Optical coherence tomographic angiography (OCTA) is a non-invasive imaging modality that provides three-dimensional, information-rich vascular images. With numerous studies demonstrating unique capabilities in biomarker quantification, diagnosis, and monitoring, OCTA technology has seen rapid adoption in research and clinical settings. The value of OCTA imaging is significantly enhanced by image analysis tools that provide rapid and accurate quantification of vascular features and pathology. Today, the most powerful image analysis methods are based on artificial intelligence (AI). While AI encompasses a large variety of techniques, machine-learning-based, and especially deep-learning-based, image analysis provides accurate measurements from a variety of contexts, including different diseases and regions of the eye. Here, we discuss the principles of both OCTA and AI that make their combination capable of answering new questions. We also review contemporary applications of AI in OCTA, which include accurate detection of pathologies such as choroidal neovascularization, precise quantification of retinal perfusion, and reliable disease diagnosis.

https://www.youtube.com/watch?v=_VOueRtahFI