Lightweight Hybrid Ensemble for Ameloblastoma and Ameloblastic Carcinoma Classification

• Domingos L. L. de Oliveira IFSP / UFU
• Hanna B. C. M. F. de Castro UNIFESP
• Thaína A. A. Tosta UNIFESP
• Leandro A. Neves UNESP
• Pedro A. A. Oliveira UFU
• Paulo R. de Fariall UFU
• Marcelo Z. do Nascimento UFU

Resumo

The accurate diagnosis of rare odontogenic tumors, such as ameloblastoma and ameloblastic carcinoma, is essential for appropriate clinical management, but is complicated by their overlapping histopathological patterns. Whole-slide imaging is essential for capturing high-resolution histological details over large tissue areas; however, it generates thousands of image patches per slide, resulting in a massive volume of data that demands efficient and scalable classification models. This study proposes a Lightweight Hybrid Ensemble (LHE) that combines three CNN-Transformer architectures (EdgeNeXt, EfficientViT, and LeViT), optimized for a low parameter count and minimal computational cost (0.8GMACs,13.9M parameters). Evaluated across four magnification levels (5×,10×,20×, and 40×), the ensemble achieved its highest performance at 40× with an accuracy of 89.0% and an F1-score of 0.89, aligning with clinical diagnostic practices that rely on high-resolution details. Analysis across multiple magnifications and external validation on an independent dataset reinforce the method's robustness. These findings highlight the potential of lightweight hybrid ensembles for reliable and efficient histopathological classification, particularly in resource-constrained settings, filling a critical gap in the literature.

Palavras-chave: Training, Solid modeling, Accuracy, Systematics, Computational modeling, Image edge detection, Performance gain, Robustness, Computational efficiency, Tumors