Point-wise spatial network for identifying carcinoma at the upper digestive and respiratory tract

Lei Zhou, Huaili Jiang, Guangyao Li, Jiaye Ding, Cuicui Lv, Maoli Duan, Wenfeng Wang, Kongyang Chen, Na Shen, Xinsheng HuangBMC Medical Imaging 2023 · DOI

TL;DR

This paper introduces a novel deep learning model, PSANet, for the unified detection of carcinomas across the oral pharynx, hypopharynx, and larynx (collectively rUDRT). By leveraging a point-wise spatial attention mechanism, the model achieves high accuracy in segmenting cancerous regions, suggesting a common visual pattern for these tumors that can be exploited by a single AI system.

What Problem Does This Paper Address?

Early detection of cancers in the upper digestive and respiratory tract (rUDRT) is crucial for improving patient prognosis. While AI has shown promise in detecting specific rUDRT cancers (e.g., oral or nasopharyngeal), a comprehensive model capable of identifying carcinomas across multiple rUDRT regions simultaneously has been lacking. This gap necessitates the development of a unified approach that can recognize common visual characteristics of tumors in these anatomically related areas, thereby facilitating broader and more efficient early diagnosis.

Methods

The core of this research is the application of a Point-wise Spatial Attention Network (PSANet) for semantic segmentation. PSANet is designed to capture long-range contextual information by allowing each point in a feature map to attend to all other points, both near and far. This is achieved through a bi-directional information propagation path that aggregates global context. The model's architecture incorporates an attention mechanism to enhance feature representation, enabling it to effectively distinguish between normal and cancerous tissues. For comparison, several classical semantic segmentation models like FCN-UNet, PSPNet, DeepLabv3, and EMANet were also evaluated. The authors also explored adding auxiliary losses to the PSANet model to potentially improve its performance.

Datasets

The study utilized a dataset of 1742 cancerous endoscopic images from 101 patients and 6473 normal or benign lesion images from 200 patients. All cancerous images were histologically confirmed squamous carcinomas. To augment the dataset and improve model robustness, various image enhancement techniques were applied, increasing the number of cancerous images to 8725. The dataset was split into training (80%), validation, and testing sets. The images were preprocessed by resizing to 480x480 pixels, converting to RGB, and standardizing pixel values.

Key Insights and Ideas

  • Unified rUDRT Cancer Detection: The research strongly suggests that carcinomas in the oral pharynx, hypopharynx, and larynx share common visual characteristics, enabling a single AI model to detect them effectively. This challenges the traditional approach of developing region-specific models.
  • Power of Spatial Attention: The PSANet's point-wise spatial attention mechanism proved highly effective in aggregating global contextual information, leading to superior performance compared to other semantic segmentation models. This highlights the importance of understanding relationships between distant image regions for accurate segmentation.
  • Potential for Self-Diagnosis: The development of an integrated AI model for rUDRT cancer detection, when combined with portable self-examination equipment, holds significant promise for facilitating early diagnosis and patient self-monitoring.

Key Results

The PSANet model demonstrated excellent performance:

  • Average mIoU: 86.3%
  • Average Pixel Accuracy: 96.3%
  • Sensitivity: 94.39%
  • Specificity: 98.68%
  • AUC (ROC curve): 0.97

These metrics significantly outperformed other tested models, with PSANet achieving the highest mIOU (86.83%) and mAcc (92.38%) in initial comparisons. The addition of auxiliary losses slightly decreased mIOU but improved overall accuracy to 96.3%.

Limitations and Future Work

The authors acknowledge several limitations:

  • Single Center Data: The study relied on images from a single tertiary care center, limiting the diversity of the dataset.
  • Limited Image Variety: The number and variety of endoscopic images were insufficient.
  • Lack of Precancerous Lesion Classification: The model was not trained to differentiate cancerous lesions from other precancerous conditions.

Future work should focus on:

  • Collecting and analyzing multicenter images.
  • Developing classification capabilities to distinguish cancer from benign tumors or infections.
  • Exploring combinations of the AI model with portable self-diagnosis devices.
  • Extending the model to other types of mucosa cancers.

Summary updated

BibTeX
@article{zhou2023point,
  author = {Lei Zhou and Huaili Jiang and Guangyao Li and Jiaye Ding and Cuicui Lv and Maoli Duan and Wenfeng Wang and Kongyang Chen and Na Shen and Xinsheng Huang},
  title = {{Point-wise spatial network for identifying carcinoma at the upper digestive and respiratory tract}},
  journal = {BMC Medical Imaging},
  year = {2023},
  doi = {10.1186/s12880-023-01076-5},
}
Figure