Enhancing precision in lymphoma detection with novel deep transformer-based neural networks.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Lymphoma presents as swollen lymph nodes and immune deficiency, often causing fatigue and weight loss.
• Computer-assisted analysis of Positron Emission Tomography (PET) scans aids in identifying metabolic changes for improved lymphoma diagnosis.
• Accurate segmentation of affected areas in PET images is crucial for effective treatment planning.

Purpose of the Study:

• To develop an Automatic Pre-Segmentation Model (APSM) for precise lymphoma identification in PET images.
• To leverage the Swin Transformer (ST) architecture for enhanced segmentation accuracy and efficiency.
• To improve the computer-assisted analysis of metabolic changes indicative of lymphoma.

Main Methods:

• Implementation of an Automatic Pre-Segmentation Model (APSM) utilizing the Swin Transformer (ST).
• Simultaneous training of the Swin Transformer for classification and identification, focusing on lymph node regions.
• The model recognizes pixel differences related to metabolic changes for accurate tissue and lymph node separation.

Main Results:

• The APSM achieved a 12.68% increase in segmentation accuracy and 13.38% improvement in precision.
• Overhead, error, and segmentation time were reduced by 12.73%, 9.27%, and 10.23%, respectively.
• The model demonstrated superior performance compared to existing segmentation methods.

Conclusions:

• The proposed APSM effectively segments lymphoma areas by analyzing pixel and feature variations.
• The Swin Transformer-based model offers a significant advancement in the accuracy and efficiency of lymphoma detection.
• This approach holds promise for improving patient outcomes through enhanced diagnostic capabilities.