COMPARISON OF CONVOLUTIONAL NEURAL NETWORK IMAGE CLASSIFICATION PERFORMANCE RELATIVE TO THE AMOUNT OF TRAINING DATA USING CARDIOMEGALY X-RAY IMAGES

Abstract

Deep learning simultaneously learns a large amount of data, correct answers (labels), and extracts and learns features from the data. When training a deep learning model, the amount of data in your dataset has a significant impact on model performance. However, in the case of medical images, it is difficult to collect a large amount of data due to problems such as personal information protection. Also, since dedicated expertise is required to build an effective dataset, the amount of data that can be obtained is limited. Data augmentation techniques have been utilized to improve performance in various medical artificial intelligence studies and are considered an essential process to improve the performance of deep learning models. Data augmentation allows you to learn from different types of data. However, the standard process for achieving good performance when using a specific size of training data is still unclear. We compared the classification performance of several convolutional neural network models using datasets with different amounts of data to identify the optimal amount of training data that ensures good performance. We attempted to quantify the optimal amount of data for each convolutional neural network model by comparing overall results using both real and augmented datasets.