BACKGROUND: Automated cell recognition from histologic images is a very complex task. Traditionally, the image is segmented by some methods chosen to suit the image type, the objects are measured, and then a classifier is used to determine cell type from the object's measurements. Different classifiers have been used with reasonable success, including neural networks working with data from morphometric analysis. METHODS: Image data of cells were input directly into neural networks to determine the feasibility of direct classification by using pixel intensity information. Several types of neural network and their ability to work with cells in a complex patterned background were assessed for a variety of images and cell types and for the accuracy of classification. RESULTS: Inflammatory cells from animal biomaterial implants in rabbit paravertebral muscle were imaged in histologic sections. Simple, three-layer, fully connected, back-propagation neural networks and four-layer networks with two layers of a shared-weights neural network were most successful at classifying the cells from the images, with 97% and 98% correct recognition rates, respectively. CONCLUSIONS: The high accuracy recognition rate shows the potential for direct classification of visual image pixel data by neural networks. Copyright 2003 Wiley-Liss, Inc.
BACKGROUND: Automated cell recognition from histologic images is a very complex task. Traditionally, the image is segmented by some methods chosen to suit the image type, the objects are measured, and then a classifier is used to determine cell type from the object's measurements. Different classifiers have been used with reasonable success, including neural networks working with data from morphometric analysis. METHODS: Image data of cells were input directly into neural networks to determine the feasibility of direct classification by using pixel intensity information. Several types of neural network and their ability to work with cells in a complex patterned background were assessed for a variety of images and cell types and for the accuracy of classification. RESULTS: Inflammatory cells from animal biomaterial implants in rabbit paravertebral muscle were imaged in histologic sections. Simple, three-layer, fully connected, back-propagation neural networks and four-layer networks with two layers of a shared-weights neural network were most successful at classifying the cells from the images, with 97% and 98% correct recognition rates, respectively. CONCLUSIONS: The high accuracy recognition rate shows the potential for direct classification of visual image pixel data by neural networks. Copyright 2003 Wiley-Liss, Inc.