Laplacian forests: semantic image segmentation by guided bagging.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Semantic segmentation of medical images is crucial for diagnosis.
• Random decision forests are a successful model but can be improved.
• Current methods lack specialization for diverse medical image datasets.

Purpose of the Study:

• To introduce a novel, efficient, and accurate technique for semantic segmentation of medical images.
• To improve upon the random decision forests model by modifying tree training.
• To enhance segmentation accuracy and training efficiency for diverse 3D CT scans.

Main Methods:

• Developed Laplacian Forests, a new technique modifying random decision forests.
• Replaced conventional bagging with guided bagging exploiting training image structure.
• Implemented automatic tree selection using learned image embedding (Laplacian eigenmap).

Main Results:

• Laplacian Forests demonstrated higher training efficiency compared to conventional decision forests.
• Achieved higher segmentation accuracy due to specialized decision trees.
• Validated on 256 manually segmented 3D CT scans with high variability.

Conclusions:

• Laplacian Forests offer a more efficient and accurate method for medical image semantic segmentation.
• The guided bagging and automatic tree selection contribute to improved performance.
• This technique is effective for segmenting diverse 3D CT scans.