CDS PhD student Weicheng Zhu presents at the 2022 MIDL conference held in Zürich, Switzerland
The research paper will advance the intersection of deep learning and medical imaging
CDS PhD student Weicheng (Jack) Zhu presented a research paper co-authored by CDS affiliated professor and NYU Langone Health Assistant Research Professor Narges Razavian and CDS Associate Professor of Data Science and Mathematics Carlos Fernandez-Granda, at the Medical Imaging with Deep Learning (MIDL) conference this July. The paper, “Interpretable Prediction of Lung Squamous Cell Carcinoma Recurrence With Self-supervised Learning,” fits MIDL’s objective of advancing the intersection of deep learning and medical imaging. This year’s conference was held in Zürich, Switzerland from July 6th through the 8th.
Advances in the field of deep learning have found computers are able to produce more reliable information from medical imaging than the human eye. As the interchange between deep learning and medical imaging grows, MIDL seeks to create a venue that promotes the exchange of ideas between the two fields of study. The annual conference connects deep learning and medical imaging researchers, clinicians, and healthcare companies for collaborative discussions.
“The most challenging aspect of this work is that clinicians and data scientists need to find common vocabulary and educate each other,” said Razavian, whose lab at the Center for Healthcare Innovation and Delivery Science (CHIDS) focuses on the intersection of machine learning, artificial intelligence, and medicine. The researcher says this exchange between the two fields takes a lot of back and forth to define problems and come up with successful solutions.
While a challenge, the intersection between the disciplines has been the most engaging part of the project for Zhu. “With the collaboration of clinicians, we obtained more knowledge on how to analyze a histopathology image,” said Zhu. “Without their expertise in the domain, we would not be able to develop the current method.” Zhu was among the first cohort of medical-track PhD students at CDS, and his current research explores machine learning for healthcare along with model interpretability. Razavian and Fernandez-Granda advise his PhD, and he previously earned his MS in Data Science at NYU after graduating with a bachelor’s in Honors Mathematics from NYU Shanghai.
The research paper presented at MIDL concerns lung squamous cell carcinoma (LSCC), slow-growing lung cancer with a high recurrence and metastasis rate. The factors that produce the recurrence and metastasis rates are currently unknown. The study focuses on the recurrence prediction of LSCC based on whole slide images (WSIs) stained with hematoxylin-eosin (H&E). Histopathological WSIs, which are examined for changes in tissues caused by disease, are generally large so they are processed as a series of smaller image tiles or “chunks.”
Standard self-supervised learning, a method of machine learning that learns from unlabeled sample data, doesn’t work well in processing these images because the image chunks look similar as a result of the single larger image they are part of. The researchers propose a two-stage conditional self-supervised learning (SSL) method that improves the models’ learning by clustering tiles with similar traits.
By identifying image clusters, they can explain the recurrence of histopathological risk factors which will help pathologists create new hypotheses for the morphological features associated with LSCC recurrence. “The field of self-supervised learning in medical imaging is new and relevant to almost all medical AI research,” said Razavian. However, it presents unique challenges that aren’t typically seen in regular computer vision research, so Razavian said she is excited their research can offer new insights and improved methods that will potentially help millions of patients.
Zhu presented their research at the talk virtually and is also hopeful the exposure will attract more attention to the field of research. “I am looking forward to hearing suggestions and questions which might help us in the next step,” said Zhu.
By Meryl Phair
