Comparação de Arquiteturas de Redes Neurais Convolucionais para Classificação de Cáries em Radiografias Panorâmicas

Abstract

Dental caries affects 2.5 billion people with untreated lesions in permanent teeth, according to the most recent WHO report published in 2022. In dental care services, its diagnosis on panoramic radiographs relies on the practitioner's visual reading, a step subject to inter-examiner variation and sensitive to workload. Convolutional neural network-based models have shown relevant performance on this task, but the literature lacks controlled comparisons between distinct architectural families on the same dataset. This work compares four architectures, Inception-v3, InceptionResNet-v2, Xception, and EfficientNetV2-S, in the binary classification of individual tooth crops extracted from panoramic radiographs. The dataset comprises 7,304 balanced images from the public InReDD and DENTEX repositories. One third was allocated to training with five-fold stratified cross-validation, and all architectures underwent fine-tuning on ImageNet-pretrained weights, while the remaining two thirds (4,870 images), kept outside the training and validation cycle, were reserved for evaluation via ensembling of the five models from each architecture. Xception achieved the highest mean F1-score in cross-validation (0.9458) and the highest precision (0.9818). Inception-v3 was the most stable across folds. InceptionResNet-v2 obtained the highest mean recall (0.9268), missing fewer lesions per fold. EfficientNetV2-S underperformed relative to the others across all metrics, with a mean accuracy of 91.37% and a recall standard deviation of 4.14 percentage points. In the inference set, the pattern observed in cross-validation was confirmed: Xception maintained the best balance between precision and recall, with an ensemble accuracy of 94.52% and an F1-score of 0.9433, while Inception-v3 recorded the lowest false positive rate, only 46 out of the 4,870 images evaluated. The three architectures from the Inception family were more robust under this data regime, and Xception showed the most suitable error profile for use in dental screening.