Applications of Deep Learning in Healthcare | Automated Analysis of Physiological Waveforms and Medical Images Event as iCalendar

(Science Event Tags, Computer Science, Seminars)

17 January 2019

12 - 1pm

Venue: Room 303.561

Location: City Campus

Host: Department of Computer Science

Speaker: Jonathan Rubin, PhD and is Senior Scientist at Philips Research North America and a Research Affiliate of the ALFA Group at MIT's Computer Science and Artificial Intelligence Laboratory

Abstract 

In this talk, I present two applications of training deep learning models applied to medical datasets:

1. Densely Connected Convolutional Networks for Detection of Atrial Fibrillation from Short Single-Lead ECG Recordings

The first work processes and analyzes electrocardiogram (ECG) waveforms to detect Atrial Fibrillation (AF). AF is the most common sustained cardiac arrhythmia and is associated with an increased risk of stroke and mortality. Timely diagnosis of AF is clinically desirable because interventions can be applied to prevent the loss of sinus rhythm. In this work, we developed deep learning based algorithms for classification of normal sinus rhythm (NSR), AF, other rhythms (O) and noise, using short-segment, single-channel ECG waveforms as input. We evaluated our models at the 2017 PhysioNet / Computing in Cardiology Challenge, where our entry achieved the third highest score (F1 = 0.82) out of 80 teams in the final phase of the challenge.

2. Ischemic Stroke Lesion Segmentation in CT Perfusion Scans using Conditional Generative Adversarial Neural Networks, Pyramid Pooling and Focal Loss

The second work uses fully convolutional neural networks (FCN) for segmenting ischemic stroke lesions in Computed Tomography (CT) perfusion images. Treatment of stroke is time sensitive and automatic segmentation methods present the possibility of accurately identifying lesions and improving treatment planning. Our FCN architecture makes use of pyramid pooling to provide global and local contextual information. To learn the varying shapes of the lesions, we train our network using focal loss, a loss function designed for the network to focus on learning the more difficult samples. We evaluated our approach at the 2018 Ischemic Stroke Lesion Segmentation (ISLES) challenge where it ranked among the top performers at the challenge conclusion. In addition, I will present recent work that improves the performance of the FCN model, using synthesized inputs created with a conditional generative adversarial network.

About our speaker

Jonathan Rubin, PhD is Senior Scientist at Philips Research North America and a Research Affiliate of the ALFA Group at MIT's Computer Science and Artificial Intelligence Laboratory. Jonathan received his PhD in Computer Science from the University of Auckland in 2013, under the supervision of Associate Professor Ian Watson. His PhD focused on the use of artificial intelligence in computer games. After his PhD, Jonathan worked as a researcher at Silicon Valley's Palo Alto Research Center before joining Philips Research in 2016. In his current work, he develops algorithms using deep learning to automatically analyze medical data for creating clinician decision support systems.