Large-scale gastric cancer screening and localization using multi-task deep neural network

Gastric cancer is one of the most common cancers, which ranks third among leading causes death. Biopsy gastric mucosa a standard procedure in screening test. However, manual pathological inspection labor-intensive and time-consuming. Besides, it challenging for an automated algorithm to locate small lesion regions gigapixel whole-slide image make decision correctly. To tackle these issues, we collected large-scale dataset with detailed region annotation designed analyzing framework consisting 3 networks could not only determine result but also present suspicious areas pathologist reference. Experiments demonstrated that our proposed achieves sensitivity 97.05% specificity 92.72% task Dice coefficient 0.8331 segmentation task. Furthermore, tested best model real-world scenario on 10,315 images from 4 medical centers.