Artificial convolution neural network for medical image pattern recognition

We have developed several training methods in conjunction with a convolution neural network for general medical image pattern recognition. An unconventional method of using rotation and shift invariance is also proposed to enhance the neural net performance. The structure of the artificial neural network is a simplified network structure of the neocognitron. Two-dimensional local connection as a group is the fundamental architecture for the signal propagation in the convolution neural network. Weighting coefficients of convolution kernels are formed by the neural network through backpropagated training for this artificial neural net. In addition, radiologists' reading procedure was modelled in order to instruct the artificial neural network to recognize the predefined image patterns and those of interest to experts. Our training techniques involve (a) radiologists' rating for each suspected image area, (b) backpropagation of generalized distribution, (c) trainer imposed functions, (d) shift and rotation invariance of diagnosis interpretation, and (e) consistency of clinical input data using appropriate background reduction functions. We have tested these methods for detecting lung nodules on chest radiographs and microcalcifications on mammograms. The performance studies have shown the potential use of this technique in a clinical environment. We also used a profile double-matching technique for initial nodule search and used a wavelet high-pass filtering technique to enhance subtle clustered microcalcifications. We set searching parameters at a highly sensitive level to identify all potential disease areas. The artificial convolution neural network acts as a final detection classifier to determine whether a disease pattern is shown on the suspected image area.