ThyFusionNet: A CNN-transformer framework with spatial aware sparse attention for multi modal thyroid disease...

Area of Science:

• Medical Image Analysis
• Artificial Intelligence in Healthcare
• Radiology

Background:

• Diagnosing complex thyroid lesions is challenging due to high incidence and intricate pathology.
• Current medical image analysis methods require enhancement for precision and robustness.

Purpose of the Study:

• To introduce ThyFusionNet, a novel deep learning architecture for enhanced thyroid disorder diagnosis.
• To leverage a large-scale multimodal dataset (ThyM3) of CT and ultrasound images.

Main Methods:

• Developed ThyFusionNet, a deep learning model with convolutional and transformer modules for feature-level fusion across modalities.
• Incorporated head-wise positional encodings and adaptive sparse attention for improved semantic alignment and spatial modeling.
• Utilized skip connections, gated-attention fusion, and an adaptive contrastive-entropy loss function.

Main Results:

• ThyFusionNet demonstrated superior accuracy, robustness, and generalization compared to existing leading methods.
• The model effectively exploits complementary information from multimodal thyroid imaging data.
• Experimental results highlight the model's potential for clinical application.

Conclusions:

• ThyFusionNet offers a significant advancement in AI-driven medical image analysis for thyroid disorders.
• The multimodal approach and novel architecture enhance diagnostic capabilities.
• The study underscores the potential for clinical deployment of advanced deep learning in radiology.