About Workshop
Aim
What Participants Will Learn
- Provide an understanding of cardiac anatomy and echocardiography fundamentals, including heart chambers, valves, LV wall segments, ultrasound principles and standard echo views.
- Introduce participants to AI applications in echocardiography and medical imaging, including CNNs, U-Net architecture, segmentation workflows and clinical image analysis.
- Equip participants with practical skills in handling DICOM echo data and cardiac imaging datasets using tools such as Weasis, pydicom, OpenCV and Google Colab.
- Enable participants to build and evaluate AI-based LV segmentation and EF quantification models using datasets such as CAMUS and EchoNet-Dynamic.
- Explore 3D ventricular reconstruction and end-to-end clinical AI pipelines, including DICOM preprocessing, segmentation, 3D mesh generation, visualization and volume reporting.
- Discuss clinical AI deployment, regulatory awareness and career readiness, including FDA SaMD considerations, model deployment with Gradio/ONNX and portfolio-building for medical imaging AI roles.
Structure
Day 1: Cardiac Morphological Anatomy, Ultrasound Physics, and Medical Image Deep Learning Foundations
- Anatomical Mapping: Left Ventricular (LV) wall segmentation utilizing the AHA 17-segment model; the critical role of standardized anatomical ground-truth labeling in training robust medical AI models.
-
Acoustic Physics & Artifacts: Ultrasound wave propagation dynamics (frequency vs. depth resolution trade-offs); phased vs. matrix array transducers and their relationship to acoustic shadowing and dropouts.
-
Standard Echocardiographic Planes: Cross-sectional analysis of Parasternal Long-Axis (PLAX), Short-Axis (PSAX), Apical 4-Chamber (A4C), and Apical 2-Chamber (A2C) views.
-
Hemodynamic Biomarkers: Quantitative estimation of Ejection Fraction ($EF$), Stroke Volume ($SV$), End-Diastolic ($EDV$), and End-Systolic ($ESV$) volumes.
-
DICOM Informatics Deep-Dive: Decoding file structures, parsing hidden metadata tags (Patient/Equipment UID), pixel array extraction, and multi-frame cine-loop data structures.
- Computer Vision Fundamentals: Supervised learning paradigms, Convolutional Neural Networks (CNNs), and the specific spatial advantages of the U-Net Encoder-Decoder architecture for pixel-level boundary detection.
Hands-on:
-
Lab 1: Navigating and parsing real patient echo DICOM arrays using Weasis PACS viewer.
-
Lab 2: Developing a Python pipeline (
pydicom+Matplotlib+OpenCV) to ingest, slice, and visualize multi-frame cine-loops.
Day 2: Advanced Deep Learning Pipelines, LV Segmentation, and Automated Volumetric Quantification
-
Endocardial Boundary Tracking: Automated tracking of the endocardial and epicardial contours; isolating challenges like low contrast-to-noise ratio, speckle noise injection, and inter-observer variability.
-
Data Augmentation for Low-Sample Medical Datasets: Domain-specific transformations including elastic deformations, random flipping, and synthetic speckle noise simulation using
AlbumentationsandMONAI. -
Loss Optimization: Balancing class imbalances using customized loss formulations:
Total Loss = Dice Loss + Binary Cross-Entropy (BCE) Loss -
Spatiotemporal Modeling: Moving beyond static frames—utilizing 3D CNNs and R(2+1)D spatio-temporal architectures for frame-to-frame tracking in dynamic EchoNet-Dynamic cine-videos.
-
Mathematical Evaluation Metrics: Quantifying model performance using spatial overlap (Dice Similarity Coefficient), boundary distance (Hausdorff Distance), and volumetric error (Mean Absolute Error for $EF$).
-
Regulatory & Bioethics: Exploring dataset bias, algorithmic fairness, and FDA Software as a Medical Device (SaMD) compliance pathways for clinical deployment.
Hands-on:
-
Lab 1: Training an end-to-end
MONAI/PyTorchU-Net model on the benchmark CAMUS dataset. -
Lab 2: Running video-based deep learning inference via EchoNet-Dynamic to automatically calculate ejection fractions.
-
Lab 3: Writing a robust custom augmentation script for dynamic cardiac ultrasound profiles.
Day 3: 3D Geometric Mesh Reconstruction, Full Deployable Pipelines, and Translational Careers
-
2D-to-3D Spatial Projection: Mathematical principles of the biplane Simpson's method, multi-planar voxel stacking, and point-cloud generation from isolated 2D semantic masks.
-
Geometric Mesh Modeling: Implementing 3D Slicer and the SlicerHeart extension to convert segmented point clouds into high-fidelity 3D surface meshes (exportable to
.STL/.OBJformats). -
Post-Processing Workflows: Mesh smoothing, decimation, and volumetric computation within professional visualization suites (
ParaViewandMeshLab). -
Pipeline Orchestration: Architecting an integrated production script:
Raw DICOM —> MONAI Inference —> VTK Surface Rendering —> Automated Volumetric Report -
Model Interoperability & Deployment: Compiling PyTorch models to ONNX Runtime for lightning-fast CPU/GPU execution; building a browser-accessible web application using Gradio; containerization using Docker.
-
Translation & Careers: Career progression tracks (Clinical AI Engineer, Medical Imaging Scientist, Regulatory Analyst) and portfolio strategy building on GitHub.
Hands-on:
-
Lab 1: Generating and optimizing a 3D Left Ventricular surface mesh using 3D Slicer and Python
VTK. -
Lab 2: Running a fully unified Python script that inputs a DICOM file and returns a complete diagnostic volume report.
-
Lab 3: Shipping a production-ready Gradio web app demo to showcase your model to clinicians.
Important Dates
Registration Ends
Workshop Dates
What You Will Gain
Outcomes
- Load, navigate and inspect a real echo DICOM file while understanding basic DICOM structure and metadata.
- Identify standard echocardiography views and explain key clinical metrics such as EF, EDV and ESV in simple language.
- Understand why U-Net is widely used for cardiac segmentation and describe its role in AI-based echo analysis.
- Train a U-Net model on the CAMUS dataset, report Dice score and compare Dice with Hausdorff distance for clinical AI evaluation.
- Run EchoNet inference, interpret predicted ejection fraction output and understand the basics of FDA SaMD requirements for AI tools.
- Produce a 3D LV mesh, run a complete DICOM-to-EF pipeline and deploy a Gradio EF prediction demo as a portfolio-ready project.
Who Should Attend
- Students: Undergraduate or graduate students in biomedical engineering, biotechnology, computer science, data science, healthcare technology, medical imaging, or related fields.
- Ph.D. Scholars/Researchers: Research scholars working in medical imaging, AI in healthcare, cardiac imaging, biomedical signal/image processing, clinical AI, or computational healthcare.
- Academicians/Faculty: Professors, lecturers, and academic researchers interested in AI-based echocardiography, cardiac image analysis, medical AI workflows, or applied healthcare analytics.
- Healthcare & Clinical Professionals: Cardiologists, radiologists, sonographers, clinical researchers, and healthcare professionals interested in understanding AI-assisted echo analysis and EF quantification.
- Industry Professionals: AI engineers, data scientists, medical imaging professionals, healthcare technology developers, regulatory professionals, and product teams working on clinical AI or digital health solutions.
Nurchu shirisha
Department of Biotechnology
Speciality: End-to-end ML pipeline design, Medical image data handling, DICOM pipeline literacy, Clinical domain knowledge, Data annotation awareness, Medical AI model training, Clinical metric reporting, Regulatory/FDA awareness, Experiment tracking (MLOps basics), Research paper reading, 3D medical visualization, Model deployment (ONNX/Gradio), Portfolio project creation, Industry landscape awareness, Docker/MLOps fundamentals