[Paper Review] OCU-Net: A Novel U-Net Architecture for Enhanced Oral Cancer Segmentation

Ahmed Albishri, Syed Jawad Hussain Shah+2 more|arXiv (Cornell University)|Oct 3, 2023

AI in cancer detectionComputer Science3 citations

TL;DR

OCU-Net is a novel U-Net architecture with a custom Channel and Spatial Attention Fusion (CSAF) module, integrated with Squeeze-and-Excitation blocks, Atrous Spatial Pyramid Pooling, residual connections, and multi-scale feature fusion, achieving state-of-the-art performance in segmenting oral cancer from H&E-stained whole-slide images on the OCDC and ORCA datasets. OCU-Net m, enhanced with a pre-trained MobileNet-V2 backbone, further improves accuracy and efficiency.

ABSTRACT

Accurate detection of oral cancer is crucial for improving patient outcomes. However, the field faces two key challenges: the scarcity of deep learning-based image segmentation research specifically targeting oral cancer and the lack of annotated data. Our study proposes OCU-Net, a pioneering U-Net image segmentation architecture exclusively designed to detect oral cancer in hematoxylin and eosin (H&E) stained image datasets. OCU-Net incorporates advanced deep learning modules, such as the Channel and Spatial Attention Fusion (CSAF) module, a novel and innovative feature that emphasizes important channel and spatial areas in H&E images while exploring contextual information. In addition, OCU-Net integrates other innovative components such as Squeeze-and-Excite (SE) attention module, Atrous Spatial Pyramid Pooling (ASPP) module, residual blocks, and multi-scale fusion. The incorporation of these modules showed superior performance for oral cancer segmentation for two datasets used in this research. Furthermore, we effectively utilized the efficient ImageNet pre-trained MobileNet-V2 model as a backbone of our OCU-Net to create OCU-Netm, an enhanced version achieving state-of-the-art results. Comprehensive evaluation demonstrates that OCU-Net and OCU-Netm outperformed existing segmentation methods, highlighting their precision in identifying cancer cells in H&E images from OCDC and ORCA datasets.

Motivation & Objective

• Address the scarcity of deep learning-based image segmentation research specifically targeting oral cancer in H&E-stained histopathological images.
• Overcome the challenge of limited annotated H&E image datasets for oral cancer by leveraging transfer learning and data augmentation.
• Develop a U-Net-based architecture with enhanced attention mechanisms to improve localization and feature representation for cancer cell detection.
• Achieve state-of-the-art segmentation performance on public oral cancer datasets (OCDC and ORCA) through architectural innovation and efficient feature learning.
• Enable earlier and more accurate detection of oral squamous cell carcinoma (OSCC) to support clinical decision-making in digital pathology.

Proposed method

• Proposed a novel Channel and Spatial Attention Fusion (CSAF) module to jointly emphasize critical spatial regions and informative channels in H&E images while capturing contextual dependencies.
• Integrated Squeeze-and-Excitation (SE) attention blocks to recalibrate channel-wise feature responses based on global context.
• Employed Atrous Spatial Pyramid Pooling (ASPP) with parallel dilated convolutions to capture multi-scale contextual information across varying receptive fields.
• Incorporated residual blocks within skip-connections to stabilize training and improve gradient flow in deep encoder-decoder architectures.
• Utilized multi-scale feature fusion to aggregate high-resolution and high-level semantic features for improved segmentation precision.
• Adopted a pre-trained ImageNet MobileNet-V2 as the encoder backbone in OCU-Net m to enhance feature extraction and model generalization under data scarcity.

Figure 1: (A) ORCA Dataset image and annotation examples: (a) Training set example. (b) Testing set example. Classes include (White: Carcinoma pixels), (Gray: Non-carcinoma tissue pixels), and (Black: Non-tissue pixels). (B) OCDC Dataset image and annotation examples: (c) Training set example. (d) T

Experimental results

Research questions

• RQ1 Can a custom attention mechanism (CSAF) improve the segmentation accuracy of oral cancer in H&E-stained histopathological images compared to standard attention modules?
• RQ2 To what extent does integrating ASPP, SE blocks, and multi-scale fusion enhance the performance of a U-Net-based model for oral cancer segmentation?
• RQ3 How does transfer learning using a pre-trained MobileNet-V2 backbone affect the segmentation performance and computational efficiency of OCU-Net on limited oral cancer datasets?
• RQ4 Does OCU-Net achieve state-of-the-art performance on benchmark oral cancer datasets (OCDC and ORCA) in terms of Dice score, IoU, and DSC metrics?
• RQ5 Can the proposed architecture generalize across diverse histological subtypes and cancer stages with limited annotated training data?

Key findings

• OCU-Net achieved superior segmentation performance on the OCDC dataset, outperforming existing methods with a Dice score of 0.892 and IoU of 0.798.
• On the ORCA dataset, OCU-Net achieved a Dice score of 0.876 and IoU of 0.781, demonstrating strong generalization across different data distributions.
• OCU-Net m, which uses a pre-trained MobileNet-V2 backbone, achieved state-of-the-art results with a Dice score of 0.901 on OCDC and 0.883 on ORCA, indicating improved performance through transfer learning.
• Ablation studies confirmed that the CSAF module contributed significantly to performance gains, with a 4.2% increase in Dice score compared to baseline U-Net without attention mechanisms.
• The integration of multi-scale fusion and residual blocks in skip-connections led to a 3.1% improvement in Dice score over standard U-Net with only encoder-decoder skip-connections.
• The model demonstrated robustness to class imbalance and small lesion detection, with high sensitivity (0.86) and specificity (0.92) on both datasets.

Figure 2: Illustration of H&E images from the OCDC dataset [ 19 ] . (a) Blue highlighted regions indicate image patches extracted for the training set. (b) Whole slide image (WSI). (c) Red highlighted regions indicate image patches extracted for the testing set.