Text-Driven Tumor Synthesis

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Pathology

Background:

• AI models struggle with diverse tumor characteristics, leading to missed detections or over-detection.
• Existing tumor synthesis methods lack control over crucial features like texture and heterogeneity.
• This limits their effectiveness in addressing AI weaknesses in medical imaging analysis.

Purpose of the Study:

• To introduce TextoMorph, a novel text-driven tumor synthesis approach.
• To enable precise control over synthetic tumor characteristics using textual descriptions and masks.
• To improve AI performance on challenging medical cases by generating targeted synthetic data.

Main Methods:

• Developed a text-driven tumor synthesis method (TextoMorph) incorporating mask control.
• Utilized text mined from radiology reports to guide tumor generation.
• Employed contrastive learning across texts and CT scans, leveraging a large corpus of reports and limited image-report pairs.
• Conducted rigorous tests to evaluate the realism and diversity of synthetic tumors.

Main Results:

• TextoMorph significantly improved AI performance in key areas: +6.5% Sensitivity in early tumor detection, +3.1% in tumor segmentation for radiotherapy (NSD), and +8.2% Sensitivity in benign vs. malignant classification.
• The approach increased variability and controllability of synthetic tumors, effectively targeting AI failure cases.
• Demonstrated realistic and diverse synthetic tumors in texture, heterogeneity, boundary, and pathology.
• Reduced dependence on scarce image-report pairs through effective use of a large report corpus.

Conclusions:

• TextoMorph offers a powerful solution for generating controllable and diverse synthetic tumors.
• This text-driven approach enhances AI diagnostic capabilities in medical imaging, particularly for challenging cases.
• The method shows promise for improving AI accuracy in tumor detection, segmentation, and classification while minimizing the need for extensive annotated datasets.