A Generative Model for Probabilistic Label Fusion of Multimodal Data

Area of Science:

• Medical Imaging
• Computational Anatomy
• Image Processing

Background:

• Multi-atlas segmentation relies on label fusion to merge results from different atlases.
• Label fusion is well-established for intramodality scenarios but less explored for multimodal data.
• Existing methods face challenges when target data modality differs from atlas modalities.

Purpose of the Study:

• To review literature on label fusion methods.
• To present an extension of a generative model-based algorithm for multimodal label fusion.
• To evaluate the performance of the proposed method against existing techniques.

Main Methods:

• A generative model exploiting voxel intensity consistency within the target scan was developed.
• The method was extended to handle multimodal target data.
• Performance was evaluated using brain MRI scans from a multiecho FLASH sequence.

Main Results:

• The proposed method achieved high segmentation accuracy (Dice 86.3% across 22 brain structures).
• The generative model-based approach outperformed majority voting, statistical-atlas-based segmentation, FreeSurfer, and an adaptive local multi-atlas method.
• The algorithm demonstrated effectiveness in multimodal and cross-modality segmentation scenarios.

Conclusions:

• The extended generative model-based label fusion method offers superior performance for multimodal medical image segmentation.
• This approach enhances the practicality of multi-atlas segmentation in diverse clinical and research settings.
• Accurate segmentation of brain structures is crucial for various neurological studies and diagnostics.