NUMERO: #20260527

Título: CROSS-DOMAIN PERFORMANCE OF RF-DETR FOR ORAL LESION DETECTION USING A PUBLIC DATASET FROM ROBOFLOW

Nome do Apresentador: Henrique Kiyoshi Aires FURUKAWA

Categoria do Trabalho: Pôster - Investigação Original (IO)

Área Temática: Estomatologia

Resumo: Objective: This study aims to develop detection algorithms to support the early identification of oral lesions using the RF-DETR, a real-time light-weight specialist detection transformer architecture for object detection.
Methods: The RF-DETR was developed using a primary retrospective clinical dataset and further assessed with an external public dataset from Roboflow. The RF-DETR model was trained for object detection using COCO-pretrained weights with an input resolution of 576×576 pixels. The training process spanned 100 epochs and utilized early stopping to prevent overfitting. To increase data variability, a data augmentation strategy was implemented
Results: The model achieved strong performance on the internal test set (mAP50 of 0.8681, precision of 0.8649, recall of 0.8205, and F1-score of 0.842). However, performance dropped significantly on the external public dataset, yielding an mAP50 of 0.3629, a precision of 0.5839, a recall of 0.3688, and an F1-score of 0.4521
Conclusion: The model successfully learns lesion morphology within its original domain but shows poor cross-domain generalization, a critical limitation for clinical use. Future work should prioritize domain adaptation and diverse, multi-center datasets to improve robustness across different clinical settings.

Autor 1:  Henrique Kiyoshi Aires FURUKAWA

E-mail 1:  [email protected]

Autor 2:   Helena QUINTELLA

E-mail 2:  [email protected]

Autor 3 :  Arnaldo Vitor Barros da SILVA

E-mail 3:  [email protected]

Autor 4:  Letícia Mello BEZINELLI

E-mail 4:  [email protected]

Autor 5:  Alan Roger dos SANTOS-SILVA

E-mail 5:  [email protected]

Autor 6:  Luiz Paulo KOWALSKI

E-mail 6:  [email protected]

Autor 7:  Anna Luíza Damaceno ARAÚJO

E-mail 7:  [email protected]