10. Application of Machine Learning to Cardiac Imaging

Name: 10. Application of Machine Learning to Cardiac Imaging
Uploaded: 2020-10-22T19:36:55.000Z
Duration: 2 h 42 min 23 s

Introduction to Guest Lectures

Overview of the Lecture Series

The lecture series consists of four guest lectures throughout the semester, with two scheduled for the upcoming week and another later in the semester.

The aim is to bring in experts who possess extensive knowledge in specific areas relevant to cardiovascular medicine.

Introduction of Professor Rahul Dail

Professor Rahul Dail is introduced as today's speaker, recognized for his significant contributions to echocardiography and machine learning applications in cardiovascular medicine.

His research background includes a focus on genotype and phenotype relationships, although today’s discussion will center on imaging techniques.

Expectations from the Lecture

Interactive Format Encouraged

Professor Dail encourages audience participation through questions and interruptions during his presentation, emphasizing an opinionated approach to the material.

Importance of Practical Application

He stresses that research should transition into practical application within clinical settings rather than remaining purely academic. This requires understanding potential barriers to implementation.

Lecture Outline

Key Topics Covered

An introduction to cardiac structure and function, which may not be familiar to all attendees at MIT. This foundational knowledge is crucial for understanding diagnostics and machine learning integration into practice.

Discussion on major cardiac diagnostics and their usage will guide thoughts on automating processes with machine learning and artificial intelligence in clinical practice.

Cardiovascular Disease Significance

Prevalence of Cardiovascular Disease

Cardiovascular disease remains the leading cause of death globally, highlighting its importance even amidst discussions about communicable diseases prevalent in developing regions. Understanding this context is vital for framing future research efforts.

Functionality of the Heart

The Anatomy and Function of the Heart

Overview of Heart Function

The heart pumps approximately five liters of blood per minute, which can increase significantly during exercise, especially in conditioned athletes.

A pause longer than three seconds can lead to lightheadedness or fainting, emphasizing the need for a consistent heartbeat throughout life.

Blood Flow Through the Heart

Blood enters the heart through the inferior and superior vena cavae, draining from the brain and lower body into the right atrium.

From the right atrium, blood flows through the tricuspid valve into the right ventricle, which pumps it to the lungs for oxygenation.

Oxygenated blood returns to the left atrium and moves into the left ventricle via the mitral valve before being pumped out to the body through the aorta.

Electrical System of the Heart

The electrical conduction system begins at the sinoatrial (SA) node in the right atrium; this is reflected in an EKG as various waves (P wave, PR interval, QRS complex).

The Wiggins diagram illustrates how electrical signals coordinate with mechanical functions like heart sounds and pressure changes during cardiac cycles.

Cardiac Cycle Dynamics

The heart undergoes cyclical phases: diastole (filling phase), systole (contraction phase), where valves open/close based on pressure changes.

This cycle repeats continuously, integrating both electrical impulses and mechanical actions for effective pumping.

Imaging Techniques in Cardiology

MRIs provide detailed images of heart anatomy but are expensive; echocardiography offers functional insights despite lower image quality.

Different imaging views help assess cardiac function and disease states by revealing specific anatomical features relevant to diagnosis.

Common Heart Diseases

Heart failure occurs when pumping efficiency declines, leading to symptoms like breathlessness; treatments include medications or devices.

Coronary artery disease results from blocked blood supply; severe cases may lead to myocardial infarction (heart attack), treated via angioplasty or bypass surgery.

Valvular Diseases and Arrhythmias

Valvular diseases arise from abnormal blood flow due to stenosis (narrowing valves) or regurgitation (leaky valves), causing symptoms like dizziness or shortness of breath.

Understanding Cardiac Physiology and Imaging

The Complexity of Heart Cells

The heart comprises various cell types, with only 30% being cardiomyocytes responsible for contraction and electrical activity.

Other significant cells include endothelial cells, fibroblasts, and numerous blood cells, indicating a complex biological environment beyond just pumping functions.

Historical Perspectives on Cardiac Disease

Traditional views of cardiac disease focus primarily on pumping efficiency and electrical activation but overlook the broader biological complexities involved.

Cost Implications of Cardiac Imaging

Cardiology heavily relies on imaging techniques which can be expensive; for instance, echocardiograms are relatively affordable compared to MRIs or angiography.

Different imaging modalities serve specific purposes: echocardiography assesses structure/function while MRI is less common due to high costs despite its detailed insights.

Non-Invasive Technologies in Cardiology

Non-invasive methods like PET scans utilize radio-nucleotides to detect blood flow abnormalities without requiring invasive procedures.

Defining Diseases Through Imaging

Current practices often define diseases based on anatomical or physiological norms derived from imaging rather than underlying biology.

Decisions regarding interventions (e.g., defibrillator placement or angioplasty) typically rely on imaging data, highlighting a reliance on existing diagnostic frameworks.

Challenges in Data Collection for Innovation

Innovating new risk models is hindered by the limited availability of comprehensive data; much of it is collected based on what insurers deem worth paying for.

Accessing Medical Imaging Data

Hospitals prioritize clinical operations over data accessibility for research purposes, complicating efforts to gather large datasets necessary for machine learning applications.

Issues with Data De-identification

Obtaining clean datasets is challenging due to identifiable information embedded in images; vendors make it difficult to de-identify this data effectively.

Understanding the Challenges of Medical Imaging

Data Integration and Scale

The integration of clinical data with imaging data is complex, as they are often stored separately. Accessing this information is crucial for comprehensive analysis.

The scale of data is significant; for instance, Brigham and Women's Hospital has about 30 million ECG records historically stored.

Imaging Modalities and Their Limitations

Different imaging modalities have unique benefits and limitations. For example, heart movement complicates imaging due to the need for high temporal frequency to avoid motion blur.

Techniques like SPECT and PET acquire images over minutes, which can be problematic for moving organs like the heart.

Image Quality Concerns

High-resolution imaging is challenging due to poor spatial resolution in certain modalities when dealing with motion.

Gating techniques are employed to align different heartbeats using ECG data, addressing issues related to image registration.

Manual Interpretation in Cardiology

Interpreting medical images often involves manual measurements by specialists using tools like calipers to assess artery narrowing or chamber sizes.

Diagnosing conditions such as cardiac amyloid relies on visual inspection, highlighting a classification challenge at both image and video levels.

Advancements in Image Analysis Techniques

Software tools have improved measurement accuracy but still require manual input from physicians, indicating room for further automation.

Key areas of focus include image classification (assigning labels to images) and semantic segmentation (classifying each pixel), though challenges remain in image registration.

The Evolution of Image Classification in Medicine

Image classification mimics physician tasks; radiologists can quickly identify conditions like lung cancer or pneumonia within minutes.

Interest in automated image classification surged post-deep learning advancements, although medicine has been slower to adopt these technologies compared to other fields.

Data Availability Challenges

Accessing large datasets is critical for effective machine learning applications; without sufficient data, progress stalls.

Historical Context of Machine Learning in Medical Imaging

The Challenges of Task Shifting in Radiology

The Role of Radiologists and Automation

The discussion begins with the limitations of automating radiological tasks, emphasizing that even if a radiologist takes only one to two minutes to read an image, their expertise is irreplaceable.

Radiologists remain crucial due to legal liabilities; they are among the most sued professionals in medicine, which discourages task shifting despite the potential for automation.

There is a reluctance among radiologists to delegate responsibilities, as they prefer to maintain control over critical decisions rather than relying on others for initial assessments.

Triage and Resource Allocation

A scenario is presented where triage could be beneficial: while a radiologist sleeps at home, automated systems might prioritize emergency room studies for review.

Despite potential efficiencies in processing images, every study would still require final review by a radiologist; only the order of assessment may change.

Visual Confirmation in Medicine

Medical practice often necessitates confirmation of visual findings. For instance, identifying a tumor requires precise localization before surgical intervention can occur.

Shared decision-making is highlighted as essential; physicians must engage patients by visually explaining conditions using comparative imaging techniques.

Predictive vs. Descriptive Accuracy

A tension exists between predictive accuracy (how well models forecast outcomes) and descriptive accuracy (how well they explain findings), complicating advancements in medical imaging technology.

The field faces challenges due to its demanding nature and inflexibility regarding integrating new technologies into established practices.

Understanding Disease Detection Models

The effectiveness of disease detection models has been questioned; there’s ongoing research into how these models can better explain classifications based on image analysis.

Two primary methods are discussed: finding exemplar images that activate class scores or analyzing specific aspects driving classifications within given images.

Advances in Image Analysis Techniques

Recent developments involve comparing intensity patterns within images to enhance classification accuracy. This approach aims to clarify which features contribute positively or negatively to diagnostic outcomes.

Understanding Medical Imaging and Algorithmic Challenges

The Role of MRI in Medicine

A brain MRI's patch size affects the resolution for localizing relevant areas, indicating a growing demand from the medical field for improved imaging techniques.

Limitations of Current Algorithms

Initial attempts to utilize algorithms for medical imaging were unsatisfactory, suggesting that advancements may have occurred since then.

Traditional Methods in Cardiology

The speaker describes their experience during a cardiology fellowship, where they manually traced circles and used rulers to compute volumes, highlighting the labor-intensive nature of current practices.

Popular Architectures in Medical Imaging

The U-Net architecture is favored for semantic segmentation in medical literature due to its encoding/decoding structure with skip connections that enhance localization.

Common Issues with Pixel-Level Classification

Problems arise when pixel-level classification leads to misidentification (e.g., satellite ventricles), which human experts would typically avoid. This highlights challenges in capturing global architectural context.

Exploring Advanced Techniques: Dilated Convolutions and Image Registration

Addressing Scale Limitations

Dilated convolutions expand receptive fields without increasing parameters, offering potential solutions to common issues faced by researchers in medical imaging.

Importance of Image Registration

Image registration is crucial for aligning scans from different cardiac cycles; however, it remains an underexplored area within computer vision despite its long-standing relevance.

Emerging Techniques: Conditional Variational Autoencoders

New methods like conditional variational autoencoders are being explored for learning geometric transformations, indicating ongoing innovation in image registration techniques.

Frustrations with Clinical Medicine and Early Disease Detection

Challenges in Cardiology Practice

The speaker expresses frustration over cardiology's slow adaptation to detect early onset diseases effectively, emphasizing missed opportunities for intervention as patients age.

Patient Experience and Treatment Delays

Understanding Optimal Values in Health Metrics

The Concept of Optimal Values

The optimal value for health metrics, such as blood pressure, serves as a reference point. For instance, optimal blood pressure is considered to be less than 120/80 mmHg.

Many individuals may present with significantly higher readings (e.g., in the 200s), while others might fall within the 140s and 150s range.

Patient Attitudes Towards Medication

Patients often exhibit a sense of nihilism regarding their health conditions, leading to excuses for delaying medication initiation.

Common reasons include personal conflicts or bad experiences that distract from addressing health issues.

Variability and Trust in Health Measurements

Inherent Variability of Risk Factors

Risk factors like blood sugar levels can show significant variability; thus, single clinic visit measurements may not be reliable indicators of overall health.

Averages over multiple visits tend to provide more accurate insights into a patient's condition but still carry biases influenced by stress and other life factors.

Consequences of Delayed Treatment

Prolonged periods without treatment can lead to rapid declines in health once symptoms manifest, particularly in conditions like heart failure.

Heart failure has a high mortality rate post-hospitalization, emphasizing the importance of early intervention before symptoms appear.

The Need for Cost-effective Monitoring Solutions

Addressing Asymptomatic Phases

There is a critical need for low-cost monitoring solutions during asymptomatic phases to prevent deterioration in patient health.

Patients desire individual-level feedback on their health metrics rather than generalized public health data; they want comparisons over time (e.g., yearly EKG or echo results).

Importance of Individualized Metrics

Effective monitoring should reflect individual progress and ideally improve with adherence to therapy.

Simple tools like ultrasounds can track changes such as left ventricular mass thickening due to high blood pressure over time.

Automating Health Data Collection

Potential for Automated Interpretation

Automated systems could enhance data collection efficiency, especially when applied early in disease progression rather than at advanced stages requiring complex decisions.

Shifting Focus to Primary Care Settings

Early Disease Management and Innovation

Introduction to Early Disease Symptoms

The early stages of the disease often present no significant symptoms, leading to decisions about medication initiation or intensification being low-risk and cost-effective.

Emphasis on focusing efforts on enhancing data collection at a lower cost rather than merely implementing existing practices.

Data Utilization and Triage Systems

The goal is to elevate medical practice by quantitatively tracking intermediate states of diseases, which can lead to better patient outcomes.

Automated systems for ECG interpretation have been in use for decades, with significant advancements noted since the early 2000s when specific patterns indicating critical conditions were identified.

Historical Context of ECG Interpretation

A quality measure established in the early 2000s mandated rapid response (within 90 minutes) upon identifying critical heart conditions through ECG readings.

Previously, delays occurred as cardiologists had to be contacted and physically present before any action could be taken, resulting in lost time during emergencies.

Advancements in Emergency Response

Automation allowed non-cardiologists like ambulance personnel to make preliminary assessments based on ECG data, expediting emergency responses without finalizing decisions prematurely.

This triage system improved patient outcomes despite potential false positives, highlighting the importance of timely intervention over waiting for specialist confirmation.

Echo Studies: Challenges and Innovations

Understanding Echo Studies

An echo study consists of multiple video recordings requiring skilled technicians for proper acquisition; this process can take up to an hour.

The complexity involved necessitates experienced personnel, which poses challenges due to their high costs and limited availability.

Data Volume and Automation Efforts

In 2011 alone, Medicare recorded seven million echo studies; thus, there exists a vast amount of archived data ripe for automation.

A recent paper aimed at automating all processes related to echo studies was published last year, emphasizing that partial automation is insufficient if human involvement remains necessary throughout other steps.

Comprehensive Automation Goals

The objective was to automate from raw data acquisition through analysis while ensuring quality control across various views used in echocardiography.

Project Development and Algorithmic Applications

Initial Project Setup

The speaker discusses the challenges of engaging graduate students in projects that apply existing algorithms, noting that while they may not be excited about the applications, they are willing to provide advisory support.

Coding and Algorithm Exploration

The speaker expresses enthusiasm for coding as a preferable alternative to homework, emphasizing the importance of finding interesting algorithmic problems to solve.

Classifying Views in Medical Imaging

The project aims to classify medical imaging views more accurately than previous publications, specifically focusing on distinguishing structures like the left ventricle and atrium when they are cut off.

Data Labeling Challenges

The speaker recounts their experience labeling data during long commutes, highlighting the difficulty of obtaining sufficient data due to the specialized nature of tasks involved in image segmentation.

Implementation of Modified Algorithms

A modified version of a unit algorithm was implemented with penalties for stray structures, allowing for comprehensive analysis across all frames in cardiac cycles rather than just two frames typically used by echo readers.

Data Acquisition and Analysis

Accessing Medical Studies

The speaker describes creating a keystroke encoder to automate downloading studies due to restrictions on bulk downloads, resulting in a collection of approximately 30,000 studies over a year.

Evaluating Measurement Accuracy

Discussion on Bland-Altman plots reveals that manual measurements often contain errors; these plots help visualize differences between automated and manual values against their means.

Critique of Correlation Metrics

The speaker references Bland's critique from 30 years ago regarding correlation coefficients being inadequate metrics for assessing diagnostic accuracy due to potential biases.

Manual vs. Automated Measurements

Understanding Automated Ejection Fraction Measurement

Challenges in Current Functionality

The process involves identifying errors in cycles and feeding them back for improvement, indicating a need for enhanced functionality.

Current comparison methods involve one person tracing two images, which is inefficient compared to processing hundreds of frames per study.

Variability and Comparison Standards

Inter-observer variability can reach 8% to 9%, while the reference standard shows up to 60%, highlighting significant discrepancies in measurement reliability.

Engaging multiple cardiologists for extensive studies is impractical; alternative comparisons with modalities like MRI are suggested but not yet implemented.

Correlation Analysis as an Approach

The team aims to find correlations within study structures, such as heart mass and pressure increases, to assess performance against existing standards.

Overall performance appears comparable to current methods, with no definitive superiority established; this reflects ongoing challenges in defining a reliable gold standard.

Innovations in Imaging Technology

Point-of-Care Imaging Solutions

There’s potential for low-cost imaging at point-of-care settings, particularly beneficial for patients undergoing chemotherapy who require regular screening echoes.

Direct Disease Detection Models

Exploring blended models that combine human expertise with biological data could enhance detection capabilities beyond traditional imaging systems.

Addressing Rare Diseases through Surveillance

Focus on Hypertrophic Cardiomyopathy

Hypertrophic cardiomyopathy is highlighted as a critical condition that often goes undetected; early identification can prevent sudden deaths among young athletes.

Development of Classification Models

Understanding Cardiac Amyloid and Valve Prolapse Detection

Overview of Cardiac Amyloid

The discussion begins with the challenges in detecting cardiac amyloid, a disease gaining attention due to new therapies. There is a noted correlation between model outputs and known factors, indicating potential for improvement in detection methods.

Model Development for Valve Prolapse

A new model was developed to analyze mitral valve prolapse using approximately 250-300 cases and several thousand controls. The focus is on identifying abnormal valve behavior during specific phases of the cardiac cycle.

Abnormalities in Valve Function

The presentation explains how a prolapsing valve appears different from a normal valve, particularly during one critical phase of the cardiac cycle where displacement is measured.

Automated Image Analysis

An automated model was created to phase images accurately, focusing on relevant parts of the cardiac cycle to enhance detection capabilities.

Video Segmentation Techniques

The approach involves segmenting video data rather than analyzing static images. This method utilizes variations in volume to determine time points within the cardiac cycle without relying on ECG signals.

Handheld Device Comparisons

Discussion shifts to handheld devices used for imaging; while they may not have as high frame rates, their image quality remains comparable to traditional systems. Training data limitations are acknowledged.

Sensitivity vs. Specificity Trade-offs

The conversation addresses balancing sensitivity and specificity when developing models for rare diseases like amyloidosis. Higher sensitivity is preferred despite potential false positives since cardiologists can quickly review flagged cases.

Cost-Benefit Analysis of Detection Models

Emphasis is placed on evaluating costs associated with false positives against benefits gained from identifying true positive cases, especially given the urgency for treatment options available for conditions like amyloidosis.

Challenges in Disease Identification

Identifying patients who could benefit from new therapies remains difficult; thus, there’s an ongoing need for effective screening tools that minimize burden on healthcare providers while maximizing patient identification success rates.

Future Considerations

Understanding Disease Detection Modalities

Importance of Multiple Modalities in Disease Detection

The speaker emphasizes the need for various modalities to detect diseases, particularly those that are sensitive enough to catch early stages.

A combination of findings from different areas (neuro, eye, cardiac) is suggested to ideally capture the disease in its most treatable state.

Patent and Academic Collaboration

UCSF is filing for a patent related to their work; however, the speaker believes their code remains freely available for academic non-profit use.

The goal is to demonstrate what is possible in scalable ways while encouraging industry collaboration for commercial product development.

Challenges in Implementation

A collaborator from New Zealand faces resource limitations with a backlog of patients and insufficient specialists, prompting interest in quick studies and automation.

The discussion shifts towards balancing accuracy goals with practical implementation challenges, questioning whether perfection should be pursued when clinician involvement will always be necessary.

Industry Adoption and Practice Change

The speaker notes that mimicking existing practices can ease adoption since changing established practices is more challenging than simply providing tools that assist clinicians.

Current ECG practices illustrate how automated measurements have become standard despite initial resistance; transformative changes require significant buy-in.

Future Directions and Innovations

Collaborative efforts are needed to develop better solutions rather than merely automating existing tasks; this requires a shift in mindset among practitioners.

There’s potential for point-of-care automated diagnoses, especially in emergency situations where rapid assessments are critical.

Addressing Skill Gaps in Data Acquisition

The necessity of skilled personnel for data acquisition poses barriers; innovations must enable less skilled individuals to perform essential tasks effectively.

Exploring New Uses for Echo in Medicine

The Challenge of Data Availability

The speaker discusses the difficulty in finding new applications for echocardiography (echo) due to a lack of existing data that demonstrates its value beyond current uses.

Emphasizes the "chicken and egg" problem, where larger datasets are needed to enable more complex studies, even if they consist of just a few video datasets.

Tracking Treatment Responses

Highlights the importance of tracking patients over time to observe treatment responses, particularly within pharmaceutical companies conducting phase two trials with limited timelines.

Suggests that introducing a streamlined data collection pipeline could facilitate frequent monitoring and potentially transformative insights into treatment responses.

Financial Backing and Implementation

Stresses the necessity for financial support to initiate these projects, which could lead to broader implementation once a viable use case is established.

Discusses the potential for surveillance systems that can identify conditions like amyloid deposits without disrupting clinical workflows or placing blame on healthcare providers.

Advancements in Risk Models

Introduces concepts around disease self-classification and risk models, noting that current methodologies are still quite basic compared to what is possible.

Points out that demonstrating how new risk models can improve therapy outcomes is crucial for gaining interest from practitioners who may be resistant to change.

Limitations of Current Research Practices

Addresses the challenge posed by high costs associated with echo procedures, which limits sample sizes in clinical trials and hampers subgroup analysis capabilities.

Concludes that research advancements often outpace practical changes in medical practice until more efficient data collection methods are developed.

The Role of Geometric Models in Cardiac Imaging

Historical Context and Challenges

Reflecting on past practices in cardiac imaging where geometric models were used to reconstruct 3D images from coarse multi-slice scans, highlighting both benefits and risks involved with such approaches.

Current Techniques and Limitations

Discusses attempts at integrating geometric priors into modern imaging techniques but notes challenges due to poor segmentation quality observed in current methods.

Mentions efforts using autoencoders as an indirect method for improving feature recognition but acknowledges limitations stemming from insufficient data availability.

Exploring Challenges in Heart Disease Research

Current Limitations in Research Architecture

The speaker discusses the difficulty in finding straightforward architectures for heart disease research, indicating a lack of accessible data and prior models.

There is mention of existing multi-scale modeling efforts from groups in Oxford and New Zealand, suggesting potential avenues for future exploration.

Advancements in Ultrasound Technology

A conversation about ultrasound technology highlights the emergence of affordable handheld devices, such as a $2,000 model with a subscription service.

The speaker notes that Philips offers an $8,000 handheld device, emphasizing the decreasing costs which could benefit resource-poor countries.

The versatility and cost-effectiveness of ultrasound imaging make it an attractive option for developing regions with limited medical resources.

Focus on Coronary Artery Disease

The speaker reflects on receiving significant funding ($85 million) to address coronary artery disease but acknowledges the challenges inherent in studying this condition due to its complexity and accessibility issues.

They express concern over the limitations of current imaging techniques and biological assays that are often too expensive or not sufficiently informative for effective research.

Data Collection Challenges

Emphasizing the need for high-dimensional data and large sample sizes, the speaker points out that many studies rely on repetitive clinical data without capturing complex biological processes adequately.

They highlight that diseases like coronary artery disease develop over long periods (10 to 20 years), complicating the introduction of new assays to demonstrate benefits effectively.

Exploring Circulating Cells as a Solution

The discussion shifts towards focusing on circulating cells as potential mediators in coronary artery disease, supported by various studies linking them to disease mechanisms.

Exploring White Blood Cell Compartment and Disease Course

Insights on White Blood Cells

The white blood cell compartment has the potential to significantly alter disease progression, supported by a substantial amount of data indicating its informative nature.

This cell type is accessible and expresses numerous genes involved in various biological processes, serving as a window into these mechanisms.

Computer Vision Approaches

The focus is on utilizing computer vision techniques for analyzing data from slides containing tens of thousands of cells per individual.

By introducing fluorescent dyes targeting different organelles, researchers aim to expand the phenotypic space through various perturbations that reveal attributes not present at baseline.

Cost-effective Research Strategies

The research approach emphasizes low-cost primary assays that are reproducible and responsive to therapy, allowing for extensive patient sampling—thousands monthly and tens of thousands annually.

Existing medical record data enables selective somatic sequencing and genome associations, enhancing the depth of analysis while maintaining affordability.

Collaboration and Data Utilization

A collaboration with MGH provides access to 3.5 million records related to self-counting and cytometer data over three years, facilitating significant event tracking.

The project also includes 13 million images from hundreds of thousands of patients' slides, which can be leveraged for transfer learning in acute heart attack patient studies.

Bridging Conventional Data with Innovative Techniques

The overarching goal is to connect existing imaging methods with low-cost serial assessments to broaden phenotypic understanding while keeping expenses manageable.