Robust Deep Learning Model for Dental Radiograph Classification: Tackling Class Imbalance in Multi-Condition Diagnostics

This paper introduces the CNN model for the classification of dental conditions from radiographic images, with particular attention to five classes including Cavity, Fillings, Impacted Tooth, Implant, and Normal. We utilized a repository containing dental radiography divided into training, validation, and testing sets. Substantial imbalance in the classes was achieved in weighted class adjustments, besides a number of data augmentations such as rotation, shift, and flipping. This model architecture uses three convolutional layers with batch normalization, max pooling, and dropout, thus enabling it to extract features while avoiding overfitting. The overall accuracy is 90%, with the model showing high precision and recall for the more frequent classes like Normal (F1-score: 0.97) and Fillings (F1-score: 0.87). The less frequent classes, such as Impacted Tooth and Cavity, have poor performance, showing potential scope for improving their performances. The confusion matrix and the classification report underlined both the strong points of the model with routine cases and its underperformances with the underrepresented categories. In the next steps, the work will focus on further enlargement of the dataset, fine-tuning of model architecture, and advanced data augmentation to improve the differentiation of conditions that resemble each other. These changes will definitely enhance the accuracy and reliability of the model by strengthening its application as a diagnostic support tool for dental radiography to effectively address clinical decisions.