Improved classifier for computer-aided polyp detection in CT colonography by nonlinear dimensionality reduction.

Area of Science:

• Medical Imaging
• Machine Learning
• Data Science

Background:

• Computer-aided detection (CAD) is feasible for polyp detection on CT scans.
• Colonic polyp candidate datasets are large-scale and high-dimensional.
• Effective dimensionality reduction is crucial for classifier performance.

Purpose of the Study:

• To propose a novel nonlinear dimensionality reduction method (DMLLE) for large-scale, high-dimensional datasets.
• To apply DMLLE to a colonic polyp dataset for improved polyp detection.
• To evaluate the performance of DMLLE in conjunction with a support vector machine (SVM) classifier.

Main Methods:

• Developed a Diffusion Map and Locally Linear Embedding (DMLLE) method for nonlinear dimensionality reduction.
• Used partial data as landmarks for initial mapping into a low-dimensional space.
• Mapped non-landmark samples based on nearest landmarks, preserving local geometry.
• Applied DMLLE to a dataset of 175,269 colonic polyp candidates with 155 features.

Main Results:

• DMLLE successfully mapped similar polyp shapes to close vicinities in the low-dimensional space.
• SVM classifier performance was significantly improved in the DMLLE-reduced space compared to original or PCA-reduced spaces.
• For 6-9 mm polyps, SVM with DMLLE achieved 83% sensitivity at 9 false positives per patient, a significant improvement over state-of-the-art methods (70% sensitivity, p<0.001).

Conclusions:

• DMLLE is an effective nonlinear dimensionality reduction technique for large-scale, high-dimensional medical imaging datasets.
• DMLLE enhances classifier performance by preserving intrinsic data geometry and reducing noisy features.
• The proposed DMLLE method offers a promising approach for improving computer-aided detection of colonic polyps.