Attention-Enhanced Multi-Task Deep Learning Model for Classification and Segmentation of Esophageal Lesions

Accurate detection and segmentation of esophageal lesions are crucial for diagnosing and treating gastrointestinal diseases. However, early detection of esophageal cancer remains challenging, contributing to a reduced five-year survival rate among patients. This paper introduces a novel multitask deep learning model for automatic diagnosis that integrates classification and segmentation tasks to assist endoscopists effectively. Our approach leverages the MobileNetV2 deep learning architecture enhanced with a mutual attention module, significantly improving the model's performance in determining the locations of esophageal lesions. Unlike traditional models, the proposed model is designed not to replace endoscopists but to empower them to correct false predictions when provided with additional Supporting Information. We evaluated the proposed model on three well-known data sets: Early Esophageal Cancer (EEC), CVC-ClinicDB, and KVASIR. The experimental results demonstrate promising performance, achieving high classification accuracies of 98.72% (F1-score: 98.08%) on CVC-ClinicDB, 98.95% (F1-score: 98.32%) on KVASIR, and 99.12% (F1-score: 99.00%) on our generated EEC data set. Compared to state-of-the-art models, our classification results show significant improvement. For the segmentation task, the model attained a Dice coefficient of 92.73% and an Intersection over Union (IoU) of 91.54%. These findings suggest that the proposed multitask deep learning model can effectively assist endoscopists in evaluating esophageal lesions, thereby alleviating their workload and enhancing diagnostic precision.