Text Embedded Swin-UMamba for DeepLesion Segmentation

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Radiology report interpretation

Background:

• Lesion segmentation on CT scans is crucial for assessing chronic diseases like lymphoma.
• Integrating large language models (LLMs) with imaging data can enhance segmentation by incorporating radiology report details.
• Current segmentation methods may not fully leverage the combined potential of imaging and textual data.

Purpose of the Study:

• To investigate the feasibility of integrating text descriptions into the Swin-UMamba architecture for improved lesion segmentation.
• To develop and evaluate a novel Text-Swin-UMamba model for combining imaging features and radiology report information.
• To assess the performance of the proposed model against existing lesion segmentation approaches.

Main Methods:

• Utilized the ULS23 DeepLesion dataset for training and testing.
• Integrated short-form descriptions from radiology reports into the Swin-UMamba architecture.
• Developed the Text-Swin-UMamba model for lesion segmentation.
• Compared the model's performance using Dice Score and Hausdorff distance metrics.

Main Results:

• Achieved a high Dice Score of 82 ± 18% and a low Hausdorff distance of 6.58 ± 10.64 pixels for lesion segmentation.
• The Text-Swin-UMamba model demonstrated a 37% improvement over the LLM-driven LanGuideMedSeg model (p < 0.001).
• Outperformed purely image-based models, xLSTM-UNet by 1.74% and nnUNet by 0.22%.

Conclusions:

• Integrating text information into the Swin-UMamba architecture is feasible and significantly improves lesion segmentation accuracy.
• The Text-Swin-UMamba model represents a state-of-the-art approach for lesion segmentation, effectively combining imaging and textual data.
• The developed model shows promise for enhancing clinical assessment of chronic diseases through more accurate automated lesion measurement.