Application Of FPGA-based Cellular Neural Network For Medical Image Segmentation

Medical image segmentation is always a difficult problem in radiological image interpretation.Most of conventional methods suffer from their computational complexity and hence can hardly be used for segmentation of high resolution images.In order to achieve efficiency in both computational complexity and accuracy,researchers try to used machine vision algorithm,in which cellular neural network(CeNN)is widely used.CeNN is considered as a powerful paradigm for embedded devices.Its analog and mix-signal hardware implementations are proved to be applicable to high-speed image processing,video analysis and medical signal processing with its efficiency and popularity limited by smaller implementation size and lower precision.Due to the large number of multiply-add operations in large-scale networks,the general CPU-based implementation is still hard to meet the computational performance requirements.Recently,digital implementations of CeNNs on FPGA have attracted researchers from both academia and industry due to its high flexibility and short time-to-market.However,most existing implementations are not well optimized to fully utilize the advantages of accelerators with unnecessary computational redundancy and insufficient parallelism.In order to achieve efficiency in both computational complexity and accuracy,we present an optimized cellular neural network(CeNN)based approach for segmentation.The approach is featured with quantization to significantly reduce the computational complexity and non-linear template for robustness.Then we applied this method for diagnosis of complete cytoreduction in mammography.We propose a multi-level optimization framework for energy efficient CeNN implementations on FPGAs.In particular,the optimization framework is featured with three level optimizations:system-,module-,and design-space-level.A parallel scheme is proposed to improve the computational performance.A data-reused optimization is adopted to enable the full potential of memory bandwidth.The we use parameter quantization and memory access optimization to eliminate unnecessary multiplications.A Roofline model based method conduct more rigorous performance optimization with limited hardware resources.Experimental results show that with the optimal configuration our framework can achieve an unit computational performance improvement of 1.34×speedup compared with existing implementations.Compared with other prior works,the proposed quantized and non-linear CeNN is able to achieve significant savings of 74%in resource overhead and 48.2%in energy consumption on FPGA while maintaining only up to 1.5%and 0.6%accuracy deviations for MLO and CC views,respectively.