Sequential and Parallel CNN Structures for the Classification of Lumbar Herniated Disc in MRI

Abstract

The purpose of the present study was to detect the lumbar herniated disc in lumbar spine MRI using Convolutional Neural Network with sequential and parallel models. We performed a CNN classification technique for detecting the normal and herniated disc using sequential (single-input) and parallel (multi-input) models while capturing the effect of dropout ratios and L2 regularizers on the overall accuracy of the model. To overcome the problems of overfitting of CNN model and to enhance the overall performance, we applied data augmentation to our dataset. After evaluating the 87 patients MRI data using sequential and parallel CNN structures, the sequential CNN structure provides higher accuracy of 99.31% (training accuracy) and 96.86% (test accuracy), and when we apply parallel CNN structure, the classification accuracy is also high i.e., 99.52% (training accuracy) and 95.38% (test accuracy). We conclude that the overall sequential and parallel CNN structures provide higher accuracy for the classification of normal or herniated disc in lumbar spine MRI, as compared to when we add dropouts and regularizers in the CNN model. The results demonstrate that our proposed CNN structures significantly outperform the state-of-the-art methods.