Automatic detection of squamous cell carcinoma metastasis in esophageal lymph nodes using semantic segmentation

Yi Pan, Zhuo Sun, Wenmiao Wang, Zhaoyang Yang, Jia Jia, Xiaolong Feng, Yaxi Wang, Qing Fang, Jiangtao Li, Hongtian Dai, Calvin Ku, Shuhao Wang, Cancheng Liu, Liyan Xue, Ning Lyu, Shuangmei ZouClinical 2020 · DOI

TL;DR

This paper introduces a novel deep learning approach for automatically detecting squamous cell carcinoma (SCC) metastasis in esophageal lymph node biopsies. By training a DeepLab v3 model on a custom annotated dataset of H&E-stained slides, they achieve high accuracy, paving the way for more efficient and reliable pathological diagnosis, especially for ESCC which is prevalent in Asia.

What Problem Does This Paper Address?

Esophageal squamous cell carcinoma (ESCC) is a major health concern, particularly in Asia, with lymph node metastasis (LNM) being a critical factor influencing patient prognosis. While automated detection of LNM has shown promise for other cancer types like breast cancer and adenocarcinomas, it has been largely unexplored for ESCC. Traditional pathological diagnosis relies on manual review of H&E-stained slides, which can be time-consuming and subject to inter-observer variability. This study aims to bridge this gap by developing an automated system for detecting SCC metastasis in esophageal lymph nodes, thereby assisting pathologists and improving diagnostic efficiency.

Methods

The researchers employed a semantic segmentation approach using a DeepLab v3 model with a ResNet-50 backbone. To address the class imbalance inherent in pathological data (where non-metastatic regions vastly outnumber metastatic ones), they utilized a focal loss function during training instead of the standard cross-entropy loss. The model was trained on custom-annotated whole slide images (WSIs) of esophageal lymph nodes. For inference, the model processes image patches to predict the probability of metastasis. The study also explored extending this model to detect metastasis in lung and throat lymph nodes without retraining, demonstrating its potential for broader application.

Datasets

  • Training Set: A custom dataset comprising 110 esophageal lymph node WSIs with metastasis and 132 WSIs without metastasis. These slides were manually annotated using an iPad-based application to delineate metastatic regions.
  • Test Set: A larger, independent cohort of 795 esophageal lymph node WSIs from 154 patients, with annotations confirmed by experienced pathologists.
  • Extension Datasets: 30 lung lymph node WSIs (9 positive, 21 negative) and 30 throat lymph node WSIs (20 positive, 10 negative) were used to test the model's generalization capabilities.

Key Insights and Ideas

  • SCC Metastasis Detection Gap: The study highlights a significant gap in automated LNM detection for squamous cell carcinomas, particularly ESCC, despite its high prevalence.
  • DeepLab v3 with Focal Loss: The combination of DeepLab v3 for its robust segmentation capabilities and focal loss for handling class imbalance proved effective for this challenging task.
  • Cross-Organ Generalization: The model's ability to detect metastasis in lung and throat lymph nodes without retraining suggests a potential for a more generalizable SCC metastasis detection framework.
  • Assisting Pathologists: The model's output, such as heatmaps highlighting suspicious regions, can serve as a valuable tool to guide pathologists, potentially reducing diagnostic time and improving accuracy.

Key Results

  • Esophageal Lymph Node Detection: The proposed DeepLab v3 model with focal loss achieved a slide-level accuracy of 94% on the main esophageal lymph node test set.
  • ROC Curve Performance: The model achieved a high Area Under the Curve (AUC) of 0.96 ± 0.01 (95% CI, 0.94-0.97) for slide-level classification, outperforming other tested models.
  • Node-Level Accuracy: At the lymph node level, the model achieved 99.2% sensitivity and 93.0% specificity, resulting in an overall node-level accuracy of 94.0%.
  • Model Extension Results:
    • Throat lymph node metastasis: 96.7% accuracy.
  • Lung lymph node metastasis: 90% accuracy.

Limitations and Future Work

  • Annotation Scope: The current model is trained to detect only metastatic regions. The authors suggest that annotating other tissue types, such as lymph node germinal centers and connective tissues, could help reduce false positives.
  • Image Quality: Some false negatives were attributed to blurred images due to out-of-focus scanning, highlighting the importance of high-quality image acquisition.
  • Generalizability: While promising, the model's performance on other organs and different types of metastases (e.g., adenocarcinoma) needs further investigation.
  • Clinical Workflow Integration: Larger-scale studies are needed to verify the impact of this digital assistance on efficiency and accuracy within actual clinical workflows.
  • External Validation: An external verification study is recommended to confirm the diagnostic performance and generalizability of the model.
  • Multi-institutional Datasets: Future research should leverage multi-institutional datasets to build more robust and widely applicable models.

Summary updated

Abstract

Esophageal squamous cell carcinoma (ESCC) is more prevalent than esophageal adenocarcinoma in Asia, especially in China, where more than half of ESCC cases occur worldwide. Many studies have reported that the automatic detection of lymph node metastasis using semantic segmentation shows good performance in breast cancer and other adenocarcinomas. However, the detection of squamous cell carcinoma metastasis in hematoxylin-eosin (H&E)-stained slides has never been reported. We collected a training set of 110 esophageal lymph node slides with metastasis and 132 lymph node slides without metastasis. An iPad-based annotation system was used to draw the contours of the cancer metastasis region. A DeepLab v3 model was trained to achieve the best fit with the training data. The learned model could estimate the probability of metastasis. To evaluate the effectiveness of the detection model of learned metastasis, we used another large cohort of clinical H&E-stained esophageal lymph node slides containing 795 esophageal lymph nodes from 154 esophageal cancer patients. The basic authenticity label for each slide was confirmed by experienced pathologists. After filtering isolated noise in the prediction, we obtained an accuracy of 94%. Furthermore, we applied the learned model to throat and lung lymph node squamous cell carcinoma metastases and achieved the following promising results: an accuracy of 96.7% in throat cancer and an accuracy of 90% in lung cancer. In this work, we organized an annotated dataset of H&E-stained esophageal lymph node and trained a deep neural network to detect lymph node metastasis in H&E-stained slides of squamous cell carcinoma automatically. Moreover, it is possible to use this model to detect lymph nodes metastasis in squamous cell carcinoma from other organs. This study directly demonstrates the potential for determining the localization of squamous cell carcinoma metastases in lymph node and assisting in pathological diagnosis.
BibTeX
@inproceedings{pan2020automatic,
  author = {Yi Pan and Zhuo Sun and Wenmiao Wang and Zhaoyang Yang and Jia Jia and Xiaolong Feng and Yaxi Wang and Qing Fang and Jiangtao Li and Hongtian Dai and Calvin Ku and Shuhao Wang and Cancheng Liu and Liyan Xue and Ning Lyu and Shuangmei Zou},
  title = {{Automatic detection of squamous cell carcinoma metastasis in esophageal lymph nodes using semantic segmentation}},
  booktitle = {Clinical},
  year = {2020},
  doi = {10.1002/ctm2.129},
}
Figure