A Substitution of Convolutional Layers by FFT Layers - A Low Computational Cost Version

Convolutional Neural Networks (CNNs) have lately gained popularity as a tool for accurately diagnosing and detecting liver problems. CNNs produced highly promising findings that were comparable to traditional detection approaches in terms of accuracy. However, one of their disadvantages is that the computing cost increases exponentially as the image size grows. We present a CNN-based technique that replaces the convolution layers with Fast Fourier Transform (FFT) layers. Instead of using the kernel to convolution each pixel in a huge ultrasound image, which has a processing overhead of O(n2). The image and kernel's FFTs are computed, then multiplied in the frequency domain with a computational cost of O(nlogn), then the result's inverse Fourier transform is computed. As a result, for forward and backward propagation in the CNN for liver steatosis classification from ultrasound pictures of benign and malignant fatty liver, we replaced convolution with FFT. To categorize 550 ultrasound fatty liver images, CNN layers achieved an accuracy of 89.82%, while our technique achieved a 4% loss in classification accuracy but an 83.3 percent reduction in computational time.