Armogida, Niccoló Giuseppe (2024) The use of machine learning algorithms to detect and classify dental caries. [Tesi di dottorato]

Documento PDF Tesi_dottorato_N.G.Armogida.pdf Visibile a [TBR] Amministratori dell'archivio Download (3MB) | Richiedi una copia

Abstract

Background: Machine learning (ML) is revolutionizing medicine and dentistry by improving diagnostic accuracy and treatment planning, especially through the analysis of medical images like X-rays and oral photographs. In dentistry, ML has shown promise in detecting dental caries more accurately than traditional methods by analyzing radiographic images and radiomic features, allowing early intervention and better patient outcomes. Therefore the aim of the present study is to develop an algorithm to detect and classify extension and depth of caries in occlusal images. Methods: Intraoral occlusal pictures were acquired using a reflex camera. Caries were segmented with LabelMe after preprocessing, labeled according to the ICDAS scale, and assessed for lesion depth using the E- D classification from endoral periapical radiography. From each segmented Region of Interest, radiomic features capturing quantitative shape information, pixel-level statistical information, and texture were extracted from each color channel of the image (RGB). Machine Learning algorithms classified healthy and decayed teeth using radiomic features. A robust pipeline included an 80:20 dataset split for training and testing, along with 5-fold cross-validation to optimize classifiers and select features before evaluating performance on an independent test set. Results: Different Machine Learning algorithms were tested for each group of features extracted from each color channel. The best classification performance metrics yielded highly relevant results, varying according to the color channel used for feature extraction and the selected features. Accuracy ranged from 88% to 95%, sensitivity and specificity values are also comparable, reaching 95% in the best case. Conclusion: Combining data from pictures and radiographs offers a promising diagnostic approach.

Downloads

Actions (login required)

Modifica documento