Deep Learning for Contrast-Enhanced CT Diagnosis of Cervical LN Metastasis in Oral Cancer

Overview

This retrospective study developed a deep learning (DL) model using Mask R-CNN to identify and classify cervical lymph nodes (LNs) metastasis in oral cancer patients from contrast-enhanced CT images. The model was trained and validated on a large dataset of 1466 cases with 11,013 labeled images, demonstrating improved accuracy and efficiency over manual radiologist assessment.

Background

Oral cancer frequently recurs due to lymph node metastasis, which is challenging to diagnose accurately using conventional imaging such as CT and MRI. Misdiagnosis can lead to undertreatment or overtreatment, adversely affecting prognosis. Deep learning, particularly convolutional neural networks like Mask R-CNN, offers potential to enhance LN metastasis detection by automating image segmentation and classification, reducing reliance on invasive procedures and radiologist workload.

Data Highlights

Key Findings

• The DL model accurately identified and segmented cervical lymph nodes on contrast-enhanced CT images across multiple anatomical levels.
• Positive and negative lymph nodes were distinguished with high precision, improving diagnostic accuracy for LN metastasis.
• Training utilized a large, well-annotated dataset with consensus labeling based on pathological confirmation and expert radiologist review.
• Mask R-CNN framework with ResNet101 backbone and feature pyramid network enabled effective multi-scale feature extraction and instance segmentation.
• The DL approach has potential to reduce radiologist workload and minimize invasive diagnostic procedures like fine-needle aspiration biopsies.

Clinical Implications

Implementing this DL model in clinical practice could enhance noninvasive detection of cervical LN metastasis in oral cancer, guiding surgical decision-making such as elective neck dissection. Improved accuracy may reduce both undertreatment and overtreatment, potentially improving patient outcomes. Additionally, automation can alleviate radiologist burden and standardize LN assessment.

Conclusion

This study demonstrates that a deep learning-based approach using Mask R-CNN can effectively identify and classify cervical lymph node metastasis in oral cancer from contrast-enhanced CT images, offering a promising tool to augment clinical diagnosis and treatment planning.

References

1. Wuhan University Hospital of Stomatology Ethics Committee 2020 -- Study Approval and Data Collection
2. He et al. 2017 -- Mask R-CNN Framework for Instance Segmentation
3. He et al. 2016 -- ResNet Deep Residual Learning for Image Recognition
4. Lin et al. 2017 -- Feature Pyramid Networks for Object Detection