Deep learning-based mediastinal lymph node assessment on PET/CT images without pixel-level annotations

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• N-staging is crucial for cancer diagnostics, assessing lymph node involvement to guide treatment.
• Manual assessment of lymph nodes on PET/CT scans is challenging due to low contrast and morphological heterogeneity.
• Current methods are time-consuming and can be subjective.

Purpose of the Study:

• To develop a deep learning algorithm for automated N-staging.
• To streamline the localization, classification, and staging of mediastinal lymph nodes.
• To enable weakly supervised training without pixel-level annotations.

Main Methods:

• Utilized atlas-to-patient registration for lymph node station localization.
• Employed weakly supervised learning with image-level labels and deduced pseudo-labels.
• Trained a deep learning model for lymph node station classification and automated N-staging.

Main Results:

• Achieved 0.88 accuracy, 0.72 sensitivity, and 0.90 specificity for lymph node station classification.
• Outperformed standard threshold-based approaches and PET lesion segmentation algorithms.
• Attained 0.63 accuracy for automatic N-staging, comparable to models trained with segmentation masks.

Conclusions:

• Dividing the N-staging problem into subtasks improves performance.
• Integrating prior knowledge (atlas registration) enhances model capabilities.
• Weakly supervised models can achieve comparable or superior performance to fully supervised methods.