Ensemble of Semantic Segmentation Models for Oral Epithelial Dysplasia Images

• Adriano B. Silva UFU
• Thaína A. A. Tosta UNIFESP
• Leandro A. Neves UNESP
• Alessandro S. Martins IFTM
• Paulo R. De Faria UFU
• Marcelo Z. Do Nascimento UFU

Resumo

Early detection of potentially malignant disorders such as oral epithelial dysplasia (OED) is important for preventing oral cancer. Semantic segmentation of nuclei in histopathological images provides relevant insights for pathologists. CNN-based methods have shown promise in improving histological lesion detection and segmentation processes, but achieving results with significant values in terms accuracy metrics remains a challenging task. This paper presents an ensemble approach to enhance the performance of semantic segmentation for nuclei in OED histopathology images. Six CNN models were employed, and their outputs were associated using three ensemble strategies: simple averaging, weighted averaging, and majority voting. To further enhance model robustness, a data augmentation stage was assessed. The proposed ensemble, combined with an image augmentation strategy, achieved accuracy and Dice coefficient values of 93.41 % and 0.88, respectively, on OED images. Analysis of the OED grades showed values ranging from 91.14% to 95.24 % and 0.87 to 0.90 for accuracy and Dice coefficient, respectively. These values show an improvement over the CNN segmentation models. The analysis of segmentation performance with the OED grade images is another significant contribution of this study that addresses a gap in the literature. A validation stage on three publicly available datasets demonstrated that our approach is on par with state-of-the-art methods.

Palavras-chave: Instance segmentation, Image segmentation, Solid modeling, Accuracy, Semantic segmentation, Data augmentation, Robustness, Lesions, Ensemble learning, Cancer