Related Experiment Videos
Massive-training artificial neural network coupled with Laplacian-eigenfunction-based dimensionality reduction for
Kenji Suzuki1, Jun Zhang, Jianwu Xu
1Department of Radiology, The University of Chicago, Chicago, IL 60637, USA. suzuki@uchicago.edu
IEEE Transactions on Medical Imaging
|June 24, 2010
Summary
This study introduces a dimension reduction method using Laplacian eigenfunctions (LAPs) for massive-training artificial neural networks (MTANNs) in CT colonography polyp detection. The LAP-MTANN significantly reduces training time while maintaining high accuracy in identifying polyps and minimizing false positives.
Area of Science:
- Medical Imaging and Artificial Intelligence
- Computer-Aided Detection (CAD) Systems
- Gastrointestinal Diagnostics
Background:
- Computer-aided detection (CAD) of polyps in CT colonography (CTC) faces challenges in reducing false-positive (FP) detections without compromising sensitivity.
- Massive-training artificial neural networks (MTANNs) are effective but computationally intensive due to high-dimensional input data.
- The lengthy training time of MTANNs hinders their efficient application in clinical settings.
Purpose of the Study:
- To propose and evaluate a dimension reduction method for MTANNs using Laplacian eigenfunctions (LAPs) to improve efficiency.
- To assess the performance of the proposed Laplacian eigenfunction-based MTANN (LAP-MTANN) in detecting polyps in CTC.
- To compare the LAP-MTANN with the original MTANN in terms of sensitivity, false-positive rate, and training time.
Main Methods:
- Developed a dimension reduction technique by employing Laplacian eigenfunctions (LAPs) to represent the dependence structures of input voxels.
- Implemented the LAP-MTANN, which uses selected LAPs as input features, reducing the dimensionality compared to raw voxels.
- Trained and tested the LAP-MTANN using a dataset of 246 CTC scans from 123 patients, including actual polyps and simulated false positives (rectal tubes).
Main Results:
- The LAP-MTANN with 20 LAPs demonstrated advantageous performance over the original MTANN with 171 inputs in initial tests.
- LAP-MTANN achieved comparable polyp detection performance to the original MTANN (95% sensitivity), with a similar false-positive rate (3.9 vs. 3.6 per patient).
- Training time was drastically reduced from 38 hours to 4 hours with the LAP-MTANN, and the area under the ROC curve was slightly higher (0.84 vs. 0.82).
Conclusions:
- The proposed LAP-MTANN effectively reduces the dimensionality of input data for MTANNs in CTC polyp detection.
- This dimension reduction significantly decreases training time without sacrificing diagnostic performance, making MTANNs more practical.
- LAP-MTANN represents a promising advancement for improving the efficiency and applicability of CAD systems in colon polyp screening.