Closed loop text guided framework for lung cancer lesion segmentation and quantification.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Accurate lung cancer lesion delineation is crucial for early detection and treatment.
• Conventional segmentation methods produce masks not aligned with clinical reporting standards.

Purpose of the Study:

• To introduce BiomedLoop, a text-guided framework for lung cancer segmentation that bridges the gap between automated analysis and clinical practice.
• To improve the alignment of segmentation outputs with radiologist language and reporting standards.

Main Methods:

• A pipeline coupling localization (Grounding DINO) and refinement (SEEM) with an Uncertainty Aware Feature Modulator.
• Converting mask geometric descriptors into pseudo text prompts for localization pathway fine-tuning.
• Generating structured reports compliant with TID 1500 specifications.

Main Results:

• BiomedLoop achieved higher Dice similarity coefficients compared to conventional CNNs and Segment Anything Model variants.
• The framework demonstrated consistently lower Hausdorff distances, indicating improved boundary delineation.
• Successful fine-tuning was achieved even on datasets lacking native radiology reports.

Conclusions:

• Systematic semantic-spatial joint modeling in BiomedLoop effectively connects segmentation with clinical utility.
• The framework offers a solution for resource-limited settings requiring accurate and clinically relevant segmentation.
• BiomedLoop facilitates routine diagnostic practice by integrating semantic descriptions with spatial quantification.