Deep learning system for lymph node quantification and metastatic cancer identification from whole-slide pathology...

Area of Science:

• Oncology
• Medical Imaging
• Computational Pathology

Background:

• Traditional lymph node metastasis diagnosis is time-consuming and labor-intensive.
• Deep learning (DL) diagnostic systems are emerging but require robust validation.
• This study addresses the need for performance verification of DL systems in identifying lymph node metastases.

Purpose of the Study:

• To develop and rigorously test a deep learning system for automated lymph node metastasis identification.
• To evaluate the system's accuracy and efficiency in a clinical context.
• To assess the impact of metastatic cancer differentiation on diagnostic performance.

Main Methods:

• A cascade DL algorithm combining Faster RCNN and DeepLab was used for lymph node region detection.
• A fusion of Xception and DenseNet-121 models extracted features for metastatic cancer identification.
• The system was trained on 921 whole-slide images and prospectively tested on 327 unlabeled images.

Main Results:

• The DL system achieved 97.13% accuracy in lymph node quantification.
• The fused Xception and DenseNet-121 model demonstrated a 93.53% Positive Predictive Value (PPV) and 97.99% Negative Predictive Value (NPV).
• Metastatic cancer differentiation level was found to influence recognition performance.

Conclusions:

• The developed DL system offers high efficiency and accuracy for lymph node metastasis diagnosis.
• This automated system has the potential to assist pathologists in preliminary screening for gastric cancer patients.
• Clinical implementation could streamline the diagnostic workflow for lymph node metastases.