Enhancing the Accuracy of Classifying Benign and Malignant Pancreatic Lesions Using the MVIT-MLKA Model: A Comprehensive Evaluation and Comparative Study

Accurate differentiation between benign and malignant pancreatic lesions is critical for patient management. This study aimed to develop and validate a novel deep learning network using baseline computed tomography images to predict benign and malignant pancreatic lesions. This retrospective study across three medical centers constituted a training cohort, an internal testing cohort, and an external validation cohorts. A novel hybrid model, Multi-Scale Large Kernel Attention with Mobile Vision Transformer (MVIT-MLKA), integrating CNN and Transformer architectures, was developed to classify pancreatic lesions. We compared the model's performance with traditional machine learning and deep learning methods. Moreover, we evaluated radiologists' diagnostic accuracy with and without the optimal model assistance.The MVIT-MLKA model demonstrated superior performance for predicting pancreatic lesions, outperforming traditional models and standard CNNs and Transformers. Radiologists assisted by the MVIT-MLKA model showed significant improvements in diagnostic performance compared to those without model assistance, with notable increases in both accuracy and sensitivity. Model interpretability was enhanced through Grad-CAM visualization, effectively highlighting key lesion areas.The MVIT-MLKA model effectively differentiates between benign and malignant pancreatic lesions, surpassing traditional methods and enhancing radiologist performance. This suggests that integrating advanced deep learning model into clinical practice has the potential to reduce diagnostic errors and optimize treatment strategies in clinical practices.