MULTI-LABEL DENTAL DISEASE DIAGNOSES USING COMPUTER VISION AND DEEP LEARNING APPROACHES

Abstract

Dental diseases like cavities, fillings, impacted teeth, and dental implants are often found in the same image, which creates a challenging multi-label classification task because of overlapping disease patterns and irrelevant background regions. Correct diagnosis of these dental diseases is crucial for better clinical examination and diagnosis. Deep learning has demonstrated success for dental image analysis in recent years. However, there are still limitations in dealing with the co-occurrence of disease and background interference, which can lead to a decrease in classification performance. This research suggests a weakly supervised disease-oriented approach to multi-label dental disease classification of panoramic dental radiographs to solve these problems. The proposed framework leverages the weakly supervised localization technique to generate multi-channel disease masks from bounding-box annotations, enabling it to locate the exact part of the disease and co-occurrence. Besides, disease specific regions are emphasized, and the disease irrelevant background regions are suppressed by blurring the background using Gaussian blur (smoothing filter), and thus the model is focused on the important diagnostic features. In addition, images were also resized to 224 x 224 resolution, which helped to maintain consistency of features and improve classification performance. The improved disease-centered images are then classified with the DenseNet121 architecture because of its capability of feature learning and efficient feature reuse. The results of the experiments prove the effectiveness of the proposed framework; training, validation and testing accuracy of 99.86%, 93.63% and 93.16% respectively are obtained. Furthermore, the proposed model obtained F1-scores of 0.78, 0.95, 0.93, and 0.93 for Cavity, Fillings, Impacted Teeth, and Implants, respectively, showing macro-average F1-score of 0.90 and a weighted-average F1-score of 0.93. The high AUC values were confirmed by ROC analysis, demonstrating good classification ability in all the disease categories. The results obtained show that the proposed weakly supervised disease-focused framework effectively boost the accuracy of multi-label dental disease classification on dental panoramic radiographs and improves the representation of disease-specific features.