Application of Artificial Neural Networks to Medical Image Pattern Recognition: Detection of Clustered Microcalcifications on Mammograms and Lung Cancer on Chest Radiographs

  • Authors:
  • Shih-Chung B. Lo;Jyh-Shyan J. Lin;Matthew T. Freedman;Seong K. Mun

  • Affiliations:
  • ISIS Center, Department of Radiology, Georgetown University Medical Center, 2115 Wisconsin Ave., Suite 603, Washington, DC 20007;ISIS Center, Department of Radiology, Georgetown University Medical Center, 2115 Wisconsin Ave., Suite 603, Washington, DC 20007;ISIS Center, Department of Radiology, Georgetown University Medical Center, 2115 Wisconsin Ave., Suite 603, Washington, DC 20007;ISIS Center, Department of Radiology, Georgetown University Medical Center, 2115 Wisconsin Ave., Suite 603, Washington, DC 20007

  • Venue:
  • Journal of VLSI Signal Processing Systems - special issue on applications of neural networks in biomedical image processing
  • Year:
  • 1998

Quantified Score

Hi-index 0.00

Visualization

Abstract

Three neural network models were employed to evaluate theirperformances in the recognition of medical image patterns associatedwith lung cancer and breast cancer in radiography. The first methodwas a pattern match neural network. The second was a conventionalbackpropagation neural network. The third method was abackpropagation trained neocognitron in which the signal propagationis operated with the convolution calculation from one layer to thenext. In the convolution neural network (CNN) experiment, severaloutput association methods and trainer imposed driving functions inconjunction with the convolution neural network are proposed forgeneral medical image pattern recognition. An unconventional methodof applying rotation and shift invariance is also used to enhance theperformance of the neural nets.We have tested these methods for the detection of microcalcificationson mammograms and lung nodules on chest radiographs. Pre-scanmethods were previously described in our early publications. Theartificial neural networks act as final detection classifiers todetermine if a disease pattern is presented on the suspected imagearea. We found that the convolution neural network, which internallyperforms feature extraction and classification, achieves the bestperformance among the three neural network models. These resultsshow that some processing associated with disease feature extractionis a necessary step before a classifier can make an accuratedetermination.