A deep-learning radiomics-based lymph node metastasis predictive model for pancreatic cancer: a diagnostic study

Area of Science:

• Oncology
• Radiology
• Medical Imaging

Background:

• Preoperative lymph node (LN) status is critical for pancreatic cancer treatment planning.
• Accurate preoperative LN staging remains a significant challenge in clinical practice.

Purpose of the Study:

• To develop and validate a novel radiomics-based model for precise preoperative lymph node status prediction in pancreatic cancer.
• To compare the performance of the developed model against traditional radiologist assessments and existing deep learning radiomics models.

Main Methods:

• A multiview-guided two-stream convolution network (MTCN) radiomics approach was employed, focusing on primary tumor and peri-tumor features.
• A modified MTCN (MTCN+) model was established incorporating clinical data (age, CA125) and radiologist judgment.
• The model's discriminative ability, survival fitting, and accuracy were evaluated across training, testing, and external validation cohorts.

Main Results:

• The MTCN+ model demonstrated superior discriminative ability and accuracy compared to the standard MTCN model and artificial intelligence models across all cohorts.
• The model showed good correlation between predicted and actual LN status for disease-free and overall survival.
• Performance was notable in patients with small primary tumors, although it was less accurate in assessing metastatic burden within lymph node-positive cases.

Conclusions:

• The novel MTCN+ model offers a significant advancement in predicting preoperative lymph node status for pancreatic cancer.
• This tool has the potential to correct approximately 40% of misdiagnoses made by radiologists.
• The MTCN+ model can enhance the precision of survival prognosis and inform treatment strategies for pancreatic cancer patients.