Scholarlink

In histological image analysis for Oral Squamous Cell Carcinoma (OSCC), stain normalization as a preprocessing step is underutilized. This study explores the impact of Reinhard’s color transfer method on the performance of deep learning models in semantic segmentation tasks. Using Hematoxylin and Eosin (H&E) stained images, we hypothesize that stain normalization enhances model robustness and serves as an effective data augmentation strategy, addressing the scarcity of OSCC datasets. We employ a lightweight UNet with an EfficientNet backbone, achieving performance comparable to ResNet-based architectures. To better capture OSCC’s complex cell structures, we integrate deformable convolutions, allowing adaptive receptive fields for improved segmentation. Results show that stain normalization and deformable convolutions significantly enhance model performance, making them effective for histological image segmentation in OSCC.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Adaptive Deep Learning for OSCC Histology: Leveraging Deformable Convolutions for Improved Segmentation

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Rupesh Mandal,
  • Prithvijit Das,
  • Hrishiraj Modi,
  • Ankit Prasad Verma,
  • Nupur Choudhury,
  • Mrinmoy Mayur Choudhury,
  • Muktanjalee Deka,
  • Jyoti Kumar Barman,
  • Gitu Das

摘要

In histological image analysis for Oral Squamous Cell Carcinoma (OSCC), stain normalization as a preprocessing step is underutilized. This study explores the impact of Reinhard’s color transfer method on the performance of deep learning models in semantic segmentation tasks. Using Hematoxylin and Eosin (H&E) stained images, we hypothesize that stain normalization enhances model robustness and serves as an effective data augmentation strategy, addressing the scarcity of OSCC datasets. We employ a lightweight UNet with an EfficientNet backbone, achieving performance comparable to ResNet-based architectures. To better capture OSCC’s complex cell structures, we integrate deformable convolutions, allowing adaptive receptive fields for improved segmentation. Results show that stain normalization and deformable convolutions significantly enhance model performance, making them effective for histological image segmentation in OSCC.