A CNN-Based Novel Approach for Classification of Sacral Hiatus with GAN-Powered Tabular Data Set

Abstract

Caudal epidural anaesthesia is usually the most well-known technique in obstetrics to deal with chronic back pain. Due to variations in the shape and size of the sacral hiatus (SH), its classification is a crucial and challenging task. Clinically, it is required in trauma, where surgeons must make fast and correct selections. Past studies have focused on morphometric and statistical analysis to classify it. Therefore, it is vital to automatically and accurately classify SH types through deep learning methods. To this end, we proposed the Multi-Task Process (MTP), a novel classification approach to classify the SH MTP that initially uses a small medical tabular data set obtained by manual feature extraction on computed tomography scans of the sacrums. Second, it augments the data set synthetically through a Generative Adversarial Network (GAN). In addition, it adapts a two-dimensional (2D) embedding algorithm to convert tabular features into images. Finally, it feeds images into Convolutional Neural Networks (CNNs). The application of MTP to six CNN models achieved remarkable classification success rates of approximately 90 % to 93 %. The proposed MTP approach eliminates the small medical tabular data problem that results in bone classification on deep models.