Multi-Task Learning With Hierarchical Guidance for Locating and Stratifying Submucosal Tumors.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Accurate localization and stratification of submucosal tumors in endoscopic ultrasound (EUS) images are crucial for preliminary cancer diagnosis.
• Challenges include poor contrast between digestive tract wall (DTW) layers and narrow anatomical structures, limiting current deep learning approaches.

Purpose of the Study:

• To develop a novel multi-task deep learning framework for simultaneous submucosal tumor localization and stratification using EUS images.
• To enhance tumor analysis by integrating digestive tract wall segmentation and leveraging hierarchical guidance and graph reasoning.

Main Methods:

• A multi-task framework was developed, incorporating simultaneous EUS image segmentation of the digestive tract wall (DTW), tumor localization, and tumor stratification.
• A hierarchical guidance module was employed to enhance feature representation for localization and stratification tasks using DTW and tumor probability maps.
• A graph reasoning module was integrated to incorporate non-local spatial relationships for improved tumor stratification.

Main Results:

• The proposed multi-task framework demonstrated significant improvements in both tumor localization and stratification accuracy on stomach-esophagus and intestinal EUS datasets.
• The method outperformed existing state-of-the-art object detection approaches in identifying and classifying submucosal tumors.
• The integration of DTW segmentation and graph reasoning contributed to a more reliable and interpretable tumor stratification model.

Conclusions:

• The developed multi-task framework effectively addresses the challenges of analyzing submucosal tumors in EUS images.
• Integrating DTW segmentation and graph reasoning enhances the performance and interpretability of deep learning models for gastrointestinal tumor diagnosis.
• This approach shows promise for improving the preliminary diagnosis of digestive tract tumors.