Medical Image Compression using Neural Network with HGAPSO Optimization

Abstract

Lossy medical image compression has become increasingly attractive due to a drastic increase in the number of images used for diagnosis and treatment. The work focused on developing a feed-forward neural network for compression of medical images with optimization of weights using hybrid genetic and particle swarm (HGAPSO) optimization technique. The neural network can achieve a better compressed & decompressed image only with the proper training and optimized weights. Training algorithms such as back-propagation algorithm (BPA) traps to local minima rather than the global one which degrades the quality of a reconstructed image. In this work, HGAPSO optimization is adopted to overcome the drawback of BPA. HGAPSO parameters are carefully chosen to have better exploitation & exploration in the search area, which avoids the algorithm from trapping to the local minima. High-quality results of Genetic Algorithm (GA) obtained using selection, crossover and mutation can provide quality guidance for PSO which improves the results of the proposed system. The performance of the proposed work is evaluated on a raw medical image database based on PSNR, MSE, and CR. The experiment is simulated for 16-4-16 neural network architecture and a compression ratio of 75 % is achieved. The results obtained indicated that with proper training PSNR could be improved by 1.98 %.