A Lightweight Method for Fast Deep Learning Inference in Medical Image Classification Based on Fourier Transform and Signal Processing

Abstract

Medical image classification often relies on Convolutional Neural Network (CNN) for its powerful ability to obtain accurate predictions. However, considering novel diseases such as COVID variants and complications, the medical and clinical field desires diagnosis that is both fast and accurate. This paper proposes a lightweight method that conducts deep learning-based classification in the Fourier domain without convolution operation and reduces the computational cost. The paper focuses specifically on pneumonia, which is a lung infection and a typical COVID complication. To achieve a decent accuracy that is comparable to the CNN performance, signal processing techniques, namely Fourier transform is utilized to extract features from the frequency domain. The proposed method uses Discrete Cosine Transform (DCT) to find the frequency domain values as well as other useful parameters. As part of the methodology, a fundamental Artificial Neural Network (ANN) is built to perform the classification task. In the meanwhile, two pre-trained CNN architectures, ResNet50V2 and VGG19, are implemented under the same environment as standards for comparison. With the same hyperparameters and training epochs, the ANN obtained a validation accuracy that is 2.35% lower than the CNNs but 15 times faster in training. The experimental result demonstrates the advantage of the proposed method in inference speed and model size, indicating that the overall objective is attained. The findings also open the possibility of generalizing such an approach for other medical diagnosis in the future.