Deep multi-task learning framework for gastrointestinal lesion-aided diagnosis and severity estimation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Accurate diagnosis and severity estimation of gastrointestinal (GI) tract lesions are critical for patient management.
• Traditional diagnostic methods struggle with inter-observer variability and lesion complexity.
• Existing deep learning models often address classification and severity estimation as separate, complicating tasks.

Purpose of the Study:

• To develop a deep multi-task learning framework for simultaneous classification and severity estimation of GI tract lesions.
• To improve diagnostic accuracy and overcome limitations of current methods.

Main Methods:

• A three-stage deep multi-task learning framework was proposed, utilizing four multi-class GI tract datasets.
• Multi-scale feature representation was achieved using convolutional vision transformer (CViT) blocks with enhanced multi-head attention.
• Shared features were extracted, concatenated, and refined with task-specific attention mechanisms for improved global and local information learning.

Main Results:

• The proposed framework demonstrated significant performance improvements in lesion diagnosis and severity estimation across multiple datasets.
• The model effectively enhanced fine-grained image features by integrating semantic information and focusing on representation subspaces.
• Validation across various datasets confirmed the model's effectiveness and robustness.

Conclusions:

• The deep multi-task learning framework offers a more accurate and unified approach to GI tract lesion diagnosis and severity estimation.
• This method addresses the limitations of separate task processing in existing deep learning models.
• The findings support the potential of this framework for clinical application in gastroenterology.