DEEP LEARNING BASED SINGLE IMAGE SUPER RESOLUTION FOR DENTAL PANORAMIC X-RAY IMAGE QUALITY ENHANCEMENT

• Main Author: Nurzaidan Widyahartono, Farhan
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/71951
• Institution: Institut Teknologi Bandung

Deep learning based single image super resolution (SISR) is currently one of the more studied and prevalent methods for image quality enhancement in the field of image processing. SISR itself has an advantage over other super resolution methods due to its minimal use of required input images and being less computationally intensive. The medical field is one field of study in which SISR has the most application potential. The main object of research in this work are dental panoramic X-ray images which are known to have limitations in quality and resolution. In this work, an implementation of Efficient Sub-Pixel Convolutional Neural Network (ESPCN) is applied for the task of SISR for dental panoramic X-ray images. The process of applying the SISR algorithm is divided into 3 stages, image pre-processing, model training, and model testing. The base model is a modified ESPCN architecture with 4 convolutional layers and RelU activation function, parameter tuning and re-training is also done during the testing process as to find the best model with optimal performance. During the training stage a dataset of panoramic dental X-ray images sourced from public domain websites is used, while for model testing the dataset is provided by the Faculty of Dentistry in Padjadjaran University. Evaluation of the model’s performance is done using Peak Signal to Noise Ratio (PSNR), Structural Similarity (SSIM), and Structural Disimilarity (DSSIM), where model predicted images are compared to a ground truth. The best model in this work achieved a mean PSNR of 38.8 dB with an average increase in image quality of 2% for the test dataset of 20 images compared to a benchmark of 38.2 dB. The images resulted from the model also achieved a SSIM score of 0.96 and DSSIM of 0.019 in comparison to the ground truth image.