Shah Syed Jawad Hussain, Albishri Ahmed, Wang Rong, Lee Yugyung
Computer Science, School of Science and Engineering, University of Missouri-Kansas City, Kansas City, USA.
Computer Science, School of Science and Engineering, University of Missouri-Kansas City, Kansas City, USA; College of Computing and Informatics, Saudi Electronic University, Riyadh, Saudi Arabia.
Comput Biol Med. 2025 May;189:109841. doi: 10.1016/j.compbiomed.2025.109841. Epub 2025 Mar 7.
Early detection of Oral Squamous Cell Carcinoma (OSCC) improves survival rates, but traditional diagnostic methods often produce inconsistent results. This study introduces the Oral Cancer Attention Network (OCANet), a U-Net-based architecture designed to enhance tumor segmentation in hematoxylin and eosin (H&E)-stained images. By integrating local and global attention mechanisms, OCANet captures complex cancerous patterns that existing deep-learning models may overlook. A Large Language Model (LLM) analyzes feature maps and Grad-CAM visualizations to improve interpretability, providing insights into the model's decision-making process.
OCANet incorporates the Channel and Spatial Attention Fusion (CSAF) module, Squeeze-and-Excitation (SE) blocks, Atrous Spatial Pyramid Pooling (ASPP), and residual connections to refine feature extraction and segmentation. The model was evaluated on the Oral Cavity-Derived Cancer (OCDC) and Oral Cancer Annotated (ORCA) datasets and the DigestPath colon tumor dataset to assess generalizability. Performance was measured using accuracy, Dice Similarity Coefficient (DSC), and mean Intersection over Union (mIoU), focusing on class-specific segmentation performance.
OCANet outperformed state-of-the-art models across all datasets. On ORCA, it achieved 90.98% accuracy, 86.14% DSC, and 77.10% mIoU. On OCDC, it reached 98.24% accuracy, 94.09% DSC, and 88.84% mIoU. On DigestPath, it demonstrated strong generalization with 84.65% DSC despite limited training data. The model showed superior carcinoma detection performance, distinguishing cancerous from non-cancerous regions with high specificity.
OCANet enhances tumor segmentation accuracy and interpretability in histopathological images by integrating advanced attention mechanisms. Combining visual and textual insights, its multimodal explainability framework improves transparency while supporting clinical decision-making. With strong generalization across datasets and computational efficiency, OCANet presents a promising tool for oral and other cancer diagnostics, particularly in resource-limited settings.
早期发现口腔鳞状细胞癌（OSCC）可提高生存率，但传统诊断方法的结果往往不一致。本研究引入了口腔癌关注网络（OCANet），这是一种基于U-Net的架构，旨在增强苏木精和伊红（H&E）染色图像中的肿瘤分割。通过整合局部和全局注意力机制，OCANet捕捉现有深度学习模型可能忽略的复杂癌变模式。一个大语言模型（LLM）分析特征图和Grad-CAM可视化结果以提高可解释性，深入了解模型的决策过程。
OCANet包含通道和空间注意力融合（CSAF）模块、挤压与激励（SE）块、空洞空间金字塔池化（ASPP）以及残差连接，以优化特征提取和分割。该模型在口腔来源癌症（OCDC）和口腔癌标注（ORCA）数据集以及DigestPath结肠肿瘤数据集上进行评估，以评估其通用性。使用准确率、骰子相似系数（DSC）和平均交并比（mIoU）来衡量性能，重点关注特定类别的分割性能。
OCANet在所有数据集上均优于现有先进模型。在ORCA上，它的准确率达到90.98%，DSC为86.14%，mIoU为77.10%。在OCDC上，其准确率达到98.24%，DSC为94.09%，mIoU为88.84%。在DigestPath上，尽管训练数据有限，但它仍以84.65%的DSC表现出很强的通用性。该模型显示出卓越的癌检测性能，能够以高特异性区分癌性区域和非癌性区域。
OCANet通过整合先进的注意力机制提高了组织病理学图像中肿瘤分割的准确性和可解释性。通过结合视觉和文本见解，其多模态可解释性框架提高了透明度，同时支持临床决策。凭借跨数据集的强大通用性和计算效率，OCANet为口腔及其他癌症诊断提供了一个有前景的工具，特别是在资源有限的环境中。
