A U-Net-Based Approach for Histological Tissue Segmentation Using RCAug Data Augmentation

• Domingos L. L. de Oliveira IFSP / UFU
• Thaína A. A. Tosta UNIFESP
• Leandro A. Neves UNESP
• Adriano B. Silva UFU
• Alessandro S. Martins IFTM
• Paulo R. de Fariall UFU
• Marcelo Z. do Nascimento UFU

Resumo

Histopathological analysis of tissue samples remains the gold standard for cancer diagnosis; however, the process is time-consuming and subject to inter- and intra-observer variability. Computer-aided diagnosis systems are crucial for mitigating these issues by automating key tasks, such as the segmentation of clinically relevant structures. However, the accuracy of these segmentation methods is often limited by data scarcity and morphological variability. To address these challenges, this study proposes a previously unexplored association of enhanced U-Net architectures with RCAug, a composite, structure-preserving data augmentation strategy. The study evaluated enhanced U-Net architectures across four public datasets, comparing the proposed method against conventional augmentation techniques and a baseline without augmentation. The results demonstrate that the RCAug-based approach is a consistent top-performer, statistically outperforming other strategies. Notably, our proposed approach achieves competitive performance, reaching Dice scores of 90.19% on the GlaS dataset and 91.78% on the OCDC dataset. Our results indicate that a sophisticated augmentation strategy is a critical and efficient pathway to high performance in histological segmentation, enabling standard architectures to achieve competitive results without requiring more complex models or extensive pre-training.

Palavras-chave: Training, Image segmentation, Accuracy, Pipelines, Computer architecture, Data augmentation, Transformers, Robustness, Standards, Cancer