Description

An onsite recording of this session will be included in the Symposium OnDemand offering. 

Robust Visual Identification of Under-resourced Dermatological Diagnoses with Classifier-Steered Background Masking

Presentation Time: 03:30 PM - 03:45 PM

Abstract Keywords: Deep Learning, Imaging Informatics, Fairness and Elimination of Bias
Primary Track: Foundations
Programmatic Theme: Public Health Informatics

Collecting images of rare dermatological diseases for machine learning detection applications is a costly, laborious task. It is difficult to collect enough images of these diagnoses to avoid the risk of low accuracy "in the wild." One of the sources of bias in these networks is irrelevant background pixel data. These pixels necessarily have no clinical significance, yet Deep Neural Networks will make weak correlations based on that information. To reduce their ability to do this, we introduce a masking augmentation algorithm, InfoMax-Cutout. It employs unsupervised Information Maximization losses to mask out background pixels. InfoMax-Cutout increased accuracy on classifying 319 diagnoses by 0.76%. These features generalized to an unseen diagnosis task (Fitzpatrick 17k), improving accuracy over a baseline by 43.3% and reducing Gini inequality by 20.9%. This approach of learning to separate out background pixels can increase accuracy in detecting diseases in Lower and Middle Income Countries.

Speaker(s):
Miguel Dominguez, PhD
VisualDx

Author(s):
Julie Ryan Wolf, PhD, MPH - University of Rochester Medical Center; Paritosh Prasad, MD, MBA, DTM&H - University of Rochester Medical Center; Wendemagegn Enbiale, MD, MPH, PhD - Bahir Dar University; Michael Gottlieb, MD - Rush University Medical Center; Carl T. Berdahl, MD, MS - Cedars-Sinai Medical Center; Art Papier, MD - VisualDx, University of Rochester Medical Center;

A Multi-Task Learning Approach for Segmentation of Breast Arterial Calcifications in Screening Mammograms

Presentation Time: 03:45 PM - 04:00 PM

Abstract Keywords: Deep Learning, Imaging Informatics, Machine Learning, Biomarkers
Primary Track: Applications

Screening mammogram is a standard imaging procedure to measure breast cancer risk among 45+ year old women. Quantifying breast arterial calcification (BAC) from screening mammograms is a non-invasive and cost-efficient approach to assess the future risk of adverse cardiovascular events among women, such as heart attack and stroke. However, segmentation of breast arterial calcification is an involved task and poses several technical challenges such as extremely small BAC finding, low breast arteries to breast area ratio in the mammogram images, tissue features such as breast folds and heterogeneous density, have very similar imaging appearance. In this work, we aim to address the shortcomings of existing SOTA methods, e.g., SCUNet, and analyze the comparative performance. Given the fact that we will not be able to simply resize mammogram to preserve the resolution, we adopted a patch-based methodology for segmentation using the original resolution which may hinder the model understanding of whole mammogram. We propose a multi-task learning approach for patch-based BAC segmentation by adding an auxiliary task of patch position prediction which force the model to learn breast anatomy to comprehend the locations where BAC will not occur, such as breast boundary. The proposed method achieves state-of-the-art performance compared to the baselines. To demonstrate the utility, we also validate our method on external data and provide survival analysis for CVD based on the BAC score and provide a comparison with CAC score.

Speaker(s):
Imon Banerjee, PhD
Arizona State U, Mayo Clinic

Author(s):
Aisha Urooj, Ph.D. - Mayo Clinic; Theo Dapamede, MD, PhD - Emory University; Bhavika Patel, MD - Mayo Clinic; William Charles O' Neill, MD - Emory Clinic - School of Medicine Faculty | Emory Physician Group Practice | Emory Healthcare Network; Hari Trivedi, MD - Emory University School of Medicine; Imon Banerjee, PhD - Arizona State U, Mayo Clinic;

Project Elucidate: Web-Based Single Cell Annotation Tool For Building Deep Segmentation Models on Stimulated Raman Histology

Presentation Time: 04:00 PM - 04:15 PM

Abstract Keywords: Imaging Informatics, Deep Learning, Machine Learning
Primary Track: Applications
Programmatic Theme: Translational Bioinformatics

Single-cell analysis of cancer histology offers crucial insights into the tumor microenvironment. In brain cancer research, an advanced imaging technique called Stimulated Raman Histology (SRH) allows for rapid digital imaging of brain tumor biopsies without requiring tissue staining. SRH microscopes combined with AI are currently being used for intraoperative tumor classification in neurosurgery. However, there is yet to be single-cell annotations for SRH. Project Elucidate proposes a collaborative, open-source web platform for building cell segmentation AI models in SRH.

Speaker(s):
Abhishek Bhattacharya, M.D.
NYU Langone

Author(s):
Abhishek Bhattacharya, M.D. - NYU Langone; Eric Landgraf, BS MS - University of Michigan Department of Neurosurgery; Cheng Jiang, B.S. M.S.E. - University of Michigan Department of Neurosurgery; Asadur Chowdhury, BS MS - University of Michigan Department of Neurosurgery; Todd Hollon, MD - University of Michigan Department of Neurosurgery;

Variogram Modeling of Spatially Variant Early Response to Concurrent Chemo- and Immunotherapy for Metastatic Non-Small Cell Lung Cancer

Presentation Time: 04:15 PM - 04:30 PM

Abstract Keywords: Clinical Decision Support, Biomarkers, Imaging Informatics, Precision Medicine, Disease Models, Machine Learning
Primary Track: Applications
Programmatic Theme: Clinical Research Informatics

Predicting response of metastatic non-small cell lung cancer (mNSCLC) to chemo-immunotherapy (chemoICI) by incorporating the spatial correlation structure of PET imaging has potential to support clinical decisions regarding patient- and lesion-level risk stratification. As a prelude to extending our previous framework, the “Voxel Forecast” multiscale regression for predicting spatially variant tumor response, we explored different variograms models of spatial correlation in the mNSCLC chemoICI response stetting.

Speaker(s):
Faisal Yaseen, PhD student in Biomedical and Health Informatics
UW, Seattle

Author(s):
Faisal Yaseen, PhD student in Biomedical and Health Informatics - UW, Seattle; Daniel S. Hippe, MS - University of Washington, Seattle, WA, USA; Parth Vijaykumar Soni, MS - University of Texas, Arlington, TX, USA; Shouyi Wang, PhD - University of Texas, Arlington, TX, USA; Chunyan Duan, PhD - Tongji University, Shanghai, China; Stephen R. Bowen, PhD - University of Washington, Seattle, WA, USA; Fred Hutchinson Cancer Center, Seattle;