Classification of urinary calculi using feed-forward neural networks
Journal
South African Journal of Chemistry
Date Issued
2006
Author(s)
Abstract
Recent studies have shown that more than 80% of the analysed samples of urinary calculi in our laboratory were mainly composed
of four types of calculi, consisting of the following substances: (1) whewellite and weddellite, (2) whewellite, weddellite and uric
acid, (3) whewellite, weddellite and struvite and (4) whewellite, weddellite and carbonate apatite. In this work the results of
classification of these types of calculi (using their infrared spectra in the region 1450–450 cm–1) by feed-forward neural networks
are presented. Genetic algorithms were used for optimization of neural networks and for selection of the spectral regions most
suitable for classification purposes. The generalization abilities of the neural networks were controlled by an early stopping
procedure. The best network architecture and the most suitable spectral regions were chosen using twentyfold cross-validation.
The cross-validation error for the real samples varies from 5.3% to 5.9% misclassifications, which makes the proposed method a
promising tool for the identification of these types of calculi.
of four types of calculi, consisting of the following substances: (1) whewellite and weddellite, (2) whewellite, weddellite and uric
acid, (3) whewellite, weddellite and struvite and (4) whewellite, weddellite and carbonate apatite. In this work the results of
classification of these types of calculi (using their infrared spectra in the region 1450–450 cm–1) by feed-forward neural networks
are presented. Genetic algorithms were used for optimization of neural networks and for selection of the spectral regions most
suitable for classification purposes. The generalization abilities of the neural networks were controlled by an early stopping
procedure. The best network architecture and the most suitable spectral regions were chosen using twentyfold cross-validation.
The cross-validation error for the real samples varies from 5.3% to 5.9% misclassifications, which makes the proposed method a
promising tool for the identification of these types of calculi.
Subjects
File(s)![Thumbnail Image]()
Loading...
Name
8CiiT-39.pdf
Size
348.4 KB
Format
Adobe PDF
Checksum
(MD5):7d9aeb0fc1263f6f5185c4b7df8f6623