Accelerating the design of probabilistic neural networks for computer aided diagnosis in Mammography, employing graphics processing units

Abstract

The aim of this study is to propose a Probabilistic Neural Network (PNN) classifier system that can operate on a consumer-level graphics processing unit (GPU) and thus, harvest its tremendous parallel computation potential in order to accelerate the training phase. Therefore, the computationally intensive training of a PNN classifier system incorporating the exhaustive search of feature combinations and the leave-one-out techniques, was effectively ported on a medium class GPU device. Programming of the GPU was accomplished by means of the CUDA framework. The proposed system was tested on a real training dataset comprising 80 patterns, each consisting of 20 textural features extracted from digital mammograms (40 normal and 40 containing micro-calcifications) by an experienced physician. The developed GPU-based classifier was trained and the required time was measured. The latter was then compared with the respective training time of the same classifier running on a typical CPU and programmed in the C programming language. According to experimental results, the proposed GPU-based classifier achieved significantly higher training speed, outperforming the CPU-based system by a factor that ranged from 10 to 75 times