Diagnosis of invasive pancreatic cancer in endoscopic ultrasound images leveraging translation models

Background: In guiding treatment decisions for pancreatic cancer, assessing vascular invasion is critical, particularly for determining resectability. Deep learning techniques have demonstrated potential for diagnosing vascular invasion through the analysis of pancreatic endoscopic ultrasound (EUS) images. However, challenges arise when dealing with multicenter data sources and imbalanced datasets, which may affect the performance of deep learning models. Method: EUS images were collected from 170 patients with pancreatic cancer diagnosed at three endoscopy centers using various ultrasound systems. To diagnose vascular invasion while mitigating data variations, feature and image translation models were utilized to effectively align the source and target domains. An image translation model was utilized in the proposed approach (multicenter transfer learning (MCTL)) by employing CycleGAN and customized weighted loss classification models. The performance was compared with those of a feature translation model (multicenter domain adaptation (MCDA)) and widely accepted baseline classification models. Result: The translation models compensated for the distinctive data-specific features and improved the models for classifying vascular invasion. Although the feature translation model proved effective, its applicability was limited across different datasets. The proposed MCTL approach showed superior classification performance with accuracy improvements of 26.79%, 67.26%, and 50.91% over the baseline model and 17.86%, 48.81%, and 42.50% over the MCDA model for the three imbalanced datasets. Significance: This study leveraged a deep learning approach for enhancing the diagnosis of vascular invasion in pancreatic cancer using EUS images from multiple centers and addressed the issue of data imbalance.