MM-WHS

MM-WHS evaluates whole-heart segmentation (WHS) methods across CT and MRI to benchmark and advance automated delineation of cardiac substructures for modeling and analysis of cardiac anatomy and function.

Key Features:

• Dataset: A standardized dataset of 120 three-dimensional cardiac images comprising 60 CT volumes and 60 MRI volumes acquired in clinical settings with manual delineations.
• Multi-modality evaluation: Direct comparative assessment across CT and MRI modalities using common data and metrics.
• Challenge setting: Conducted in conjunction with MICCAI 2017 to provide a formal benchmarking framework.
• Benchmark submissions: Evaluations included algorithms submitted by twelve groups, with ten CT-focused and eleven MRI-focused methods.
• Evaluation protocols: Use of standardized evaluation metrics and blinded evaluations to assess performance and generalizability.
• Training data and priors: Emphasis on initial annotated training datasets for constructing priors or training models.
• Algorithm classes: Inclusion of deep learning (DL)-based methods with varying network structures and training strategies alongside conventional multi-atlas segmentation approaches.
• Performance observations: Reported generally higher segmentation accuracy on CT than MRI and wide variability in DL performance depending on network architecture and training strategy.
• Computational considerations: Multi-atlas methods provided reliable baselines but exhibited limitations in accuracy and computational efficiency.
• Ongoing benchmarking resource: Continued provision of annotated training data and blinded evaluations to support method comparison and development.

Scientific Applications:

• Whole-heart segmentation: Delineation of cardiac substructures for quantitative anatomy extraction.
• Cardiac modeling and analysis: Construction of models to study cardiac anatomy and function from CT and MRI.
• Algorithm development and benchmarking: Training, validating, and comparing automated segmentation algorithms including deep learning and multi-atlas methods.
• Modality comparison studies: Investigation of modality-specific challenges and accuracy differences between CT and MRI.

Methodology:

Explicit methodologies include use of standardized annotated training datasets to construct priors or train models, application of deep learning architectures with varied network structures and training strategies, multi-atlas segmentation approaches, and blinded evaluations with standardized metrics.