Center for Data Science Members Present at MIDL Montréal 2020
We’re excited to announce that our PhD student Nan Wu alongside CDS affiliated postdoctoral researcher Stanisław Jastrzębski, NYU members Linda Moy and Jungkyu Park, and CDS professor Kyunghyun Cho and CDS affiliated professor Krzysztof J. Geras from the NYU Grossman School of Medicine, presented at MIDL Montréal 2020 Conference this week. The conference took place July 6–9 and was available via live stream on YouTube in Montréal time (ET = UTC-4).
The MIDL conference (acronym stands for Medical Imaging with Deep Learning) “aims to be a forum for deep learning researchers, clinicians and health-care companies to take a leap in the application of deep learning based automatic image analysis in disease screening, diagnosis, prognosis, treatment selection and treatment” — MIDL website. The three-day conference features keynote presentations from invited speakers, posters, oral presentations, and live demonstrations of deep learning algorithms from both academia and industry.
Their paper being presented, which is titled, “Improving the Ability of Deep Networks to Use Information From Multiple Views in Breast Cancer Screening”, looks at breast cancer screening with multiple views: Mediolateral Oblique (MLO) and Craniocaudal (CC). Using both views is essential to make an accurate diagnosis in breast cancer screenings. Their research question being asked: when it comes to multiview deep neural networks, is information utilized in both views and how to encourage the network to achieve better performance from this perspective?
The paper was selected as a spotlight — top 32% of accepted papers — at MIDL this year. The abstract is as follows:
In breast cancer screening, radiologists make the diagnosis based on images that are taken from two angles. Inspired by this, we seek to improve the performance of deep neural networks applied to this task by encouraging the model to use information from both views of the breast. First, we take a closer look at the training process and observe an imbalance between learning from the two views. In particular, we observe that parameters of the layers processing one of the views have larger gradient norms and contribute more to the overall loss reduction. Next, we test several methods targeted at utilizing both views more equally in training. We find that using the same weights to process both views, or using a technique called modality dropout, leads to a boost in performance. Looking forward, our results indicate improving learning dynamics as a promising avenue for improving utilization of multiple views in deep neural networks for medical diagnosis.
Here is a bit of information on each CDS participants’ background (in order of involvement.)
Nan Wu is a CDS Ph.D. student. At CDS, Nan works with Krzysztof J. Geras and Kyunghyun Cho on deep learning research for medical imaging. Currently, their research centers on both application and theoretical learning problems related with breast cancer screening. Nan holds a Master’s in Data Science from CDS.
Stanisław Jastrzębski is a postdoc at the Department of Radiology at NYU School of Medicine and an affiliated postdoc researcher at CDS and works closely with Kyunghyun Cho and Krzysztof Geras. Stanisław is also an advisor at Molecule.one. He holds a PhD in Machine Learning.
Kyunghyun Cho is a professor of Data Science at CDS and associate professor of computer science at NYU’s Courant Institute. Kyunghyun is also a CIFAR Associate Fellow and was a research scientist at Facebook AI Research from June 2017 to May 2020. He holds a Master’s in Machine Learning and Data Mining and a PhD in Machine Learning (Deep Learning.)
Krzysztof J. Geras is an assistant professor of radiology at the NYU School of Medicine and former CDS postdoctoral researcher. His core research interests include unsupervised learning with neural networks, model compression, transfer learning and evaluation of machine learning models. Krzysztof holds a MSc in Computer Science and a PhD in Informatics.
For more information on the MIDL Montréal Conference, please visit their website.
By Ashley C. McDonald