Deep Supervised Domain Adaptation for Invasive Pancreatic Cancer Diagnosis from Endoscopic Ultrasound Images

Objectives: Vascular invasion is a crucial parameter for assessing the respectability of pancreatic cancer, making accurate diagnosis essential for effective treatment decisions. Deep learning techniques have shown promise in various applications with pancreatic endoscopic ultrasound (EUS) images. However, no previous study has focused on vascular invasion in the context of pancreatic cancer. There exist domain differences among ultrasound imaging equipment and also variations in preferred imaging settings among hospitals and medical professionals. Utilizing images obtained from different domains for training deep learning models can potentially result in poor diagnostic performance. Thus, it is required to overcome domain variations in EUS images to enhance diagnostic accuracy of the deep learning model for invasive pancreatic cancer. Materials and Methods: A total of 459 EUS images were acquired from 76 patients with pancreatic cancer diagnosed from two different hospitals. To overcome the differences between domains, pre-processing methods, such as alpha-blending and multi-operator transformation, were applied and their data representations were visualized. To reduce the domain variations in the deep learning model, domain adaptation was applied. The images were categorized into source and target groups based on their respective hospitals. The performance of the proposed model was evaluated and compared with well-established models including VGG16, EfficientNetB5, and ViT-L-32. Evaluation metrics were the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and accuracy. Result: Proposed model with pre-processing and domain adaptation shows significant improvement in the diagnosis of invasive pancreatic cancer. For the source domain, the proposed model achieves an impressive 92.3% accuracy, 96.5% AUC score, 85.7% sensitivity, and 96.0% specificity. For the target domain, the model exhibits performance with 73.2% accuracy, 87.8% AUC score, 92.8% sensitivity, and 66.7% specificity. The findings validate the effectiveness of the proposed model in diagnosing invasive pancreatic cancer across domain variations.