Page History

Bridging the gap between prostate radiology and pathology through machine learning

Mirabela Rusu, PhD, Stanford University

The subtle difference in MRI appearance of prostate cancer and benign prostate tissue renders the interpretation of prostate MRI challenging, causing many false positives, false negatives, and wide variations in interpretation. My laboratory focuses on improving the interpretation of prostate MRI by developing deep learning models that automatically localize indolent and aggressive prostate cancers on MRI scans. The novelty of our methods comes from using whole-mount pathology images to label MRI images and to create pathomic MRI biomarkers of aggressive and indolent cancers. Our approach achieved an area under the receiver operator characteristics curve of 0.93 evaluated on a per-lesion basis and outperformed existing deep learning models. In patients outside our training cohorts, such predictive models will outline the extent of cancer on radiology images in the absence of pathology images, thus helping guide the prostate biopsy and local treatment.

The talk will focus on discussing recent contributions from my lab on registering whole-mount pathology images with MRI, training deep learning models to extract pathomic MRI biomarkers and using them in training deep learning models to detect and distinguish indolent and aggressive prostate cancers on MRI, and showing the benefits of using labels from pathology in training deep learning models to distinguish indolent from aggressive prostate cancer on MRI.

MP4 file

Transcript

• Machine Intelligence/Data Science in Medical Imaging of Breast Cancer and COVID-19
  (Maryellen Giger, Ph.D, University of Chicago)

  • Artificial Intelligence in medical imaging involves research in task-based discovery, predictive modeling, and robust clinical translation.  Quantitative radiomic analyses, an extension of computer-aided detection (CADe) and computer-aided diagnosis (CADx) methods, are yielding novel image-based tumor characteristics, i.e., signatures that may ultimately contribute to the design of patient-specific cancer diagnostics and treatments. Beyond human-engineered features, deep convolutional neural networks (CNN) are being investigated in the diagnosis of disease on radiography, ultrasound, and MRI.  The method of extracting characteristic radiomic features of a lesion and/or background can be referred to as “virtual biopsies”.  Various AI methods are evolving as aids to radiologists as a second reader or a concurrent reader, or as a primary autonomous reader.  In addition, performance evaluations, as well as considerations of robustness and repeatability, are necessary to enable translation. This presentation will discuss the development, validation, database needs, and ultimate future implementation of AI in the clinical radiology workflow including examples from breast cancer and COVID-19.  In addition, aspects of MIDRC (midrc.org) will be discussed.

MP4 file

Transcript

• AI/ML Trends in Oncology and the Rugged Path Towards the Clinic
  (Issam El Naqa, PhD, Moffit Cancer Center)
  Artificial intelligence (AI) and Machine learning (ML) algorithms are currently transforming biomedical research, especially in the context of cancer research and clinical care. Despite the tremendous potentials in automating workflow, personalizing care, and reducing health disparity, to name a few prospects, their application in oncology and healthcare has been limited in scope with less than 5% of major healthcare providers implementing any form of AI/ML solutions. This can be attributed to multitude of concerning issues regarding the deployment of AI/ML driven technologies into the clinic. These concerns include but not limited to skepticism related to commercialization hype, under representative training data, inherent implementation bias, lack of robustness and absence of prediction transparency. In this work, we will discuss some of these impending challenges and highlight different approaches for detecting and mitigating such bias in implementing clinical AI/ML algorithms. We further show examples of applying these approaches in oncology applications from our work and others and discuss their implications to pave the way for AI/ML in clinical practice. 

MP4 file

Transcript

• Digital twins for oncology via imaging-based mathematical modeling
  (Thomas Yankeelov, PhD, University of Texas Austin)
  Our lab is focused on integrating quantitative imaging data with mechanism-based, mathematical models to predict treatment response.  In this presentation, we will discuss some of our preliminary efforts at building digital twins to achieve this goal.  We will begin by considering the I-SPY trials for breast cancer as a specific example of how the success of adaptive, population-based clinical trials indicate that digital twins can lead to the success of adaptive, individual-based, clinical trials.  Then we will emphasize the importance of physics and biology-based mathematical models for constructing digital twins. Finally, we will illustrate how these ideas are beginning to play out in predicting and optimizing neoadjuvant therapy for breast cancer.

MP4 file

Transcript

Precision Medicine Approach to Breast Cancer Detection and Diagnosis

(Martin Yaffe, PhD, Sunnybrook Research Institute and The University of Toronto)

Dr. Yaffe will describe a multi-platform approach under investigation in his lab to improve the effectiveness of breast cancer detection and diagnosis. We are developing radiomic tools to guide the stratification of women for breast cancer screening that will be more accurate and efficient in detection than the “one size fits all” use of mammography whose accuracy suffers, particularly in dense breasts.  We employ microsimulation modeling to guide that work. We are also exploring the integration of radiomic information from in vivo medical images with histopathology, single-cell multiplex biomarker analysis, and targeted molecular sequencing to better characterize breast and other cancers and their immune environment and to explore their spatial heterogeneity.

MP4 file

Transcript (vtt)

Transcript (txt)

NCI CCR Artificial Intelligence Resource: Recent AI Applications in Cancer Imaging

Presenters:
G. Thomas Brown, MD PhD, Staff Clinician, NCI/CCR
Stephanie Harmon, PhD, Staff Scientist, NCI/CCR
Nathan Lay, PhD, Staff Scientist, NCI/CCR

Artificial Intelligence (AI) is becoming important for cancer research but is difficult to access for most labs. In 2020, the NCI Center for Cancer Research (CCR) created a new AI Resource (AI) to benefit researchers in the CCR. The group focuses on translational computer vision approaches to analyzing medical images, such as radiologic, digital pathology, video/endoscopy and optical imaging, among others.  Examples of potential projects include developing better screening, detection methods or predictive markers, or improving procedures among many others.

With experts in pathology, medical imaging, and machine learning, AIR has taken on a diverse portfolio of research projects in their first year. In this seminar, senior members of the group will discuss its formation, collaboration experience, recent progress, and challenges for deploying developed models back to the hands of researchers across varying domains in NCI.

The presentation contained unpublished data.

As projects are finished and code is released, the AIR team will update the webpage.

https://ostr.ccr.cancer.gov/emerging-technologies/air/

You can also email the AIR team with any questions you might have (air@nih.gov)

• Multiplexed Tissue Imaging to Study Cancer
  (Sandro Santagata, MD PhD)

Our team at the Harvard Medical School’s Lab of Systems Pharmacology has generated reagents, workflows, and data analysis/visualization approaches for multiplexed tissue imaging. We developed tissue-based cyclic immunofluorescence (t-CyCIF) for subcellular imaging of formalin-fixed and paraffin-embedded (FFPE) and frozen tissues across 20-60 different proteins markers from a single tissue section. To support the use of multiplexed tissue imaging in the NCI Human Tumor Atlas Network (HTAN), we have developed algorithms and workflows to analyze these complex images, digital docents for their narrated viewing, and reporting standards for public data sharing. The information from these imaging methods complement data acquired by microregion spatial transcriptomics technologies. We have also used high-resolution imaging of tissues to identify functional interactions (e.g., immune synapses) in cancer tissues and have created multiplexed 3D cancer atlases to more completely characterize the architecture of the tumor-immune landscape in colon cancer and in melanomas, from pre-cancer lesions through metastasis.

• Imaging Data Commons Production Release Update
  (Andrey Fedorov, PhD)

The National Cancer Institute (NCI) Cancer Research DataCommons (CRDC) aims to establish a national cloud-based data science infrastructure. The goal of IDC is to enable a broad spectrum of cancer researchers, with and without imaging expertise, to easily access and explore the value of deidentified imaging data and to support integrated analyses with non-imaging data. We achieve this goal by co-locating versatile imaging collections with cloud-based computing resources and data exploration, visualization, and analysis tools. The IDC pilot was released in October 2020. In this presentation, we will give a brief overview of the capabilities of the production release of the IDC platform, and discuss the next steps for the development.

MP 4 file

Transcript

• Medical Imaging De-Identification Initiative (MIDI)
  
  
• A DICOM dataset for evaluation of medical image de-identification
  • Dr. Fred Prior (UAMS)

  • A growing number of tools and procedures claim to properly de-identify image data. Based on our decade of experience managing the Cancer Imaging Archive (TCIA) on behalf of NCI, we developed a DICOM dataset that can be used to evaluate the performance of de-identification algorithms. DICOM objects were selected from datasets published in TCIA.  Synthetic Protected Health Information (PHI) was generated and inserted into selected DICOM Attributes to mimic typical clinical imaging exams. The DICOM Standard and TCIA curation audit logs guided the insertion of synthetic PHI into standard and non-standard DICOM data elements. An answer key was created based on our knowledge of the placement of synthetic data and the DICOM standard’s guidelines for what actions should be taken in regard to the synthetic PHI.  A TCIA curation team tested the utility of the evaluation dataset and answer key.
    
    

• Medical Image De-Identification using Cloud Services
  • Dina Mikdadi, Dr. Benjamin Kopchick

  • Patient privacy rules require the removal of Protected Health Information (PHI) before sharing images publicly. Manual de-identification is no longer scalable due to the rapid increase in imaging data volume. Our goal was to configure and test the efficacy of a cloud service for automated medical image de-identification (MIDI).  One such service for DICOM images is Google Cloud Platform’s Healthcare API. Training and test datasets for validation of image de-identification, specifically prepared by the placement of synthetic PHI in DICOM headers and image pixel data, were obtained from The Cancer Imaging Archive. The customized MIDI pipeline correctly performed 99.8% of expected actions on DICOM header data elements. For image pixel data, one false-positive case was noted, while all sensitive information was correctly removed from image pixel data. Throughput averaged at 58.4 images per second. This implementation of the MIDI pipeline holds promise for automated de-identification at scale. However, verification by a human expert is still currently recommended.

MP4 file

Transcript

Establishing next generation dynamic susceptibility contrast MRI based biomarkers for neuro-oncologic applications

Dr. Chad Quarles, PhD

Dynamic susceptibility contrast (DSC) MRI is one of the most widely used physiologic imaging techniques in neuro-oncology, enabling the differentiation of glioma grades, identification of tumor components in non-enhancing glioma, reliable detection of recurrence, and early detection of therapy response. This presentation will highlight how an improved understanding of the biophysics of the contrast mechanisms underlying DSC-MRI enabled the recent standardization of acquisition protocols for multi-site clinical trials, is leading to the field’s first benchmark for software validation, and is informing the development of advanced pulse sequences and analysis strategies tailored to specific clinical challenges faced in the management of brain cancer patients. 

MP4 file

Transcript

• Can you sue an algorithm?
  • Dr. Saurabh Jha (University of Pennsylvania, Perelman School of Medicine)

MP4 file

Transcript

• Precision Surgery: Intraoperative molecular imaging to improve margin detection

  Dr. Eben Rosenthal, Professor of Otolaryngology, Head & Neck Surgery and Radiology, Stanford University
  

Abstract:

Cancer is nearly always a surgically treated disease. Almost 80% of patients with early stage solid tumors undergo surgery at some point within their treatment course. A major gap in quality of care remains the high rate of tumor-positive margins in head and neck cancer (HNC) following surgical resections. Positive margin rates are directly correlated with lower survival but have remained unchanged at 25% for the last two decades! Primary factors that have impeded improving the rate of tumor-positive margins include subjective surgeon assessment as well as the limited amount of the tissue that can be sampled for intraoperative frozen-section analysis. We have demonstrated that use of intraoperative molecular imaging (IMI) can objectively identify the area on the tumor specimen most likely to contain a tumor-positive margin (“sentinel margin”). In a prospective evaluation, a fluorescently-labeled tumor-specific contrast agent is administered intravenously to the patient several days prior to surgery. After the surgical resection, the specimen is evaluated with IMI, in which near infrared imaging is used to identify the location of the sentinel margin on the surgical specimen. This evaluation is compared to subjective assessments of the deep tumor margin by palpation, considered the standard of care. It is expected that IMI imaging will be more accurate in identifying the sentinel margin, and will shorten the time to histological diagnosis while maintaining tissue orientation and high histological image quality. The translation of these new technologies has the potential to double the five-year survival rate of patients with HNC as well offer the potential to improve care for other cancer types as well. 

MP4 file

Transcript

Orchestration of distributed image archives
(Jonas Almeida and Praphulla Bhawsar)

• Summary: Recent advances in the use of HTTP range requests to traverse bioformats*, coupled with a general move to “zero footprint” image informatics solutions, enable the creation of image archives as an exercise in governance. A particular feature of this configuration is that the images do not have to be copied or moved from their primary location. This has two interesting effects: a) the image owner remains in control of its governance, and b) training of AI classifiers can be federated across image sets that are not even shared.

  * Bremer E, Saltz J, Almeida JS. ImageBox 2 – Efficient and rapid access of image tiles from whole-slide images using serverless HTTP range requests. J Pathol Inform 2020;11:29

MP4 file

Transcript

Slides

The NIAID Office of Cyber Infrastructure and Computational Biology will share their TB Portals Program, including their efforts at imaging data collection, data dissemination, tool development, and data science research. https://tbportals.niaid.nih.gov

• Program introduction – Alex Rosenthal
• Demo #1 – Alyssa Long
• Demo #2 – Andrei Gabrielian
• Imaging data science research – Ziv Yaniv & Gabriel Rosenfeld

MP4 file

Transcript

• Imaging Data Commons (Andrey Fedorov, Dennis Bontempi)

MP4 file

Transcript

Slides

• MONAI (Stephen Aylward, Prerna Dogra, Jorge Cardoso)

  • Introduction to MONAI and the MONAI Community - Stephen Aylward

  • MONAI medical deep learning capabilities and roadmap - Jorge Cardoso 

  • MONAI and clinical workflows: Clara and federated learning - Brad Genereaux

• TCIA Update (Justin Kirby)

MP4 file

Transcript

• Kheops (Joël Spaltenstein, Osman Ratib)
• TCIA Update (Justin Kirby)

MP4 file

Transcript

• Computational Imaging for Precision Medicine: A quest for generalizable AI models  (Satish Viswanath)

• TCIA Update (John Freymann)

MP4 file

Transcript

• Distributed Learning of Deep Learning in Medical Imaging (Daniel Rubin)
• MedICI website (Benjamin Bearce)
• TCIA update (John Freymann)

• ACR's AI-LAB (Laura Coombs, Chris Treml)
• TCIA update (Justin Kirby)

• PathPresenter - a web-based digital pathology and image viewer (Rajendra Singh, Matthew Hanna)
• TCIA update (Justin Kirby)

• Medical Segmentation Decathlon: Generalizable 3D Semantic Segmentation (Amber Simpson)
• Imaging Data Commons Update (Todd Pihl)
• TCIA Update (Justin Kirby)

• Data Commons Overview (Todd Pihl)
• The Imaging Data Commons (Andrey Fedorov)
• TCIA Update (Justin Kirby)
• NBIA Update (Scott Gustafson)

• HistoQC: An Open-Source Quality Control Tool for Digital Pathology Slides (Andrew Janowczyk)
• RIL-Contour: a Medical Imaging Dataset Annotation Tool for and with Deep Learning (Kenneth Philbrick)
• TCIA Update (Justin Kirby)

• Advanced Methods in Tissue Cytometry (Rupert Ecker)
• Presentation by the 4D Necleome Imaging Working Group (David Grünwald)
• TCIA Update (Justin Kirby)

Joint Session with the CPTAC Special Interest Group

• CPTAC Project Overview (Chris Kinsinger)
• CPTAC Image Data at TCIA (Justin Kirby)
• CPTAC Proteomics Data at the Proteomics Data Commons (R. Rajesh Thangudu)
• CPTAC Genomic Data at the Genomics Data Commons (Ana Robles)
• Using the CPTAC Data Portal (R. Rajesh Thangudu)