Classification of Computed Tomography Images in Different Slice Positions Using Deep Learning.

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Radiology

Background:

• Accurate classification of medical images is crucial for diagnosis.
• The impact of dataset size on deep learning model performance in CT image classification requires further investigation.

Purpose of the Study:

• To investigate the relationship between the number of computed tomography (CT) images and the accuracy of classification models.
• To compare the performance of AlexNet and GoogLeNet convolutional neural network (CNN) architectures for CT image classification.

Main Methods:

• A dataset of 1539 patients' CT images (contrast and noncontrast) was used.
• Datasets were created with varying image counts (0.1K to 10K) across 10 anatomical classes.
• AlexNet and GoogLeNet CNN models were trained and evaluated for image classification tasks.

Main Results:

• GoogLeNet achieved the best overall accuracy of 0.721 for 10-class classification using the 10K dataset.
• AlexNet achieved the best accuracy of 0.862 for slice position classification without contrast media using the 2K dataset.
• AlexNet demonstrated faster training times compared to GoogLeNet.

Conclusions:

• Increasing the number of CT images generally improves classification model accuracy.
• The choice of CNN architecture (AlexNet vs. GoogLeNet) impacts performance based on the specific classification task.
• Dataset size and model architecture are key factors in optimizing CT image classification performance.