DEEP LEARNING-BASED SEGMENTATION OF DENTAL INFECTIONS IN RADIOGRAPHIC IMAGES USING U-NET

Abstract

Clinical importance: Well-defined borders are essential in the diagnosis and treatment planning of dentition and are important for follow-up evaluation in dental X-rays. Manual segmentation is time consuming, and may be variable between observers. Objective: To build a deep learning pipeline for binary, pixel-wise dental infection region segmentation in X-rays with U-Net. The project was based on an existing annotated collection of grayscale dental X-ray images, image level annotations and binary segmentation. The data was divided in to 70% training data, 15% validation data and 15% test data. The U-Net model with 4 encoder blocks and 4 decoder blocks was implemented, along with skip connections, ReLU activation, batch normalization and dropout. The training was done on Google Colaboratory with a learning rate of 0.0001, a batch size of 16, a mixture of objectives (Dice loss and binary cross-entropy), and a planned number of 50–100 epochs with early stopping on validation loss. Results: The project documentation was archived and includes details about the model implementation, training, testing, and generation of segmentation masks separating infected and normal tissue. The source thesis, however, wasn't able to provide the number of images in the dataset, the resolution of each image, exact training time, number of outputs per image, number of images per confusion-matrix cell, and numerical dice, IoU, precision, recall, and F1. Therefore, this article is a development and proof-of-concept evaluation and is not intended to provide a diagnostic-accuracy estimate. Conclusion: The pipeline implementation shows the potential of using a compact U-Net to automate the segmentation of dental infections and suggests that comprehensive documentation of datasets and numerical result of the test sets along with external clinical validation is required prior to any diagnostic claim.