Physician-guided deep learning model for assessing thymic epithelial tumor volume.

Area of Science:

• Oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Current Response Evaluation Criteria in Solid Tumors (RECIST) uses one-dimensional measurements, which are insufficient for complex tumor morphologies.
• Thymic epithelial tumors (TETs) often present with curvilinear shapes, particularly when metastasized to the pleura, challenging accurate RECIST assessment.
• Accurate volumetric assessment is crucial for evaluating treatment response in TETs.

Purpose of the Study:

• To develop and validate a physician-guided deep learning model for efficient and reproducible volumetric assessment of thymic epithelial tumors (TETs).
• To improve the accuracy of tumor response evaluation in patients with advanced or metastatic TETs.

Main Methods:

• A retrospective study utilizing 231 CT scans from 81 patients with TETs.
• Development of a ground truth by manual outlining of tumors on CT scans.
• Quantification of artificial intelligence (AI) model performance using Dice Similarity Coefficient (DSC), absolute volume difference, and relative volume difference on an independent test set.

Main Results:

• The AI model achieved an overall DSC of 0.77 per scan when guided by physician-identified tumor bounding boxes.
• Mean absolute volume difference between AI and ground truth measurements was 16.1 cm³.
• Mean relative volume difference was 22%, demonstrating the model's accuracy in volumetric assessment.

Conclusions:

• A robust annotation workflow and AI segmentation model for advanced TETs have been successfully developed.
• The AI model offers efficient and reproducible volumetric assessments, enhancing outcome evaluations.
• Integration into Picture Archiving and Communication Systems alongside RECIST measurements can improve patient care for metastatic TETs.