TLDR: Researchers developed Interactive-MEN-RT, a specialized AI tool for 3D meningioma segmentation in radiotherapy planning. It integrates various clinician interaction methods (points, boxes, lassos, scribbles) and significantly outperforms general-purpose tools, achieving high accuracy (up to 77.6% Dice similarity) on MRI scans. This domain-specific approach promises to improve precision and efficiency in cancer treatment.
Meningiomas, the most common primary intracranial tumors, often require radiotherapy (RT) as a critical treatment. The success of this treatment hinges on the precise delivery of radiation to the tumor while sparing surrounding healthy brain tissues. However, accurately outlining these tumors from medical images, particularly Magnetic Resonance Imaging (MRI) scans, is a complex and labor-intensive task. Manual segmentation is prone to inconsistencies even among experienced clinicians due to the varied shapes, locations, and imaging characteristics of meningiomas.
While automated deep learning methods have shown promise in medical image segmentation, they often fall short in consistently achieving the high accuracy required for critical clinical applications like RT planning, especially with diverse tumor appearances. Generic interactive segmentation tools, which combine AI with clinician input, also frequently lack the specific precision needed for disease-specific tasks.
To address these significant challenges, researchers from Seoul National University of Medicine and Seoul National University Hospital have introduced Interactive-MEN-RT, a specialized Interactive Medical Image Segmentation (IMIS) tool. This innovative framework is specifically designed to assist clinicians in the 3D segmentation of meningiomas for radiotherapy planning. The goal is to enhance both the accuracy and efficiency of tumor delineation, directly impacting treatment efficacy and patient safety.
Interactive-MEN-RT integrates several clinically relevant interaction methods, making it highly adaptable to clinician preferences and diverse tumor characteristics. These methods include point annotations for localized corrections, bounding boxes to define broader regions of interest, lasso tools for precise delineation of irregular tumor boundaries, and scribbles for indicating larger areas for inclusion or exclusion. This flexibility is crucial for real-world clinical scenarios.
The system is built upon the robust nnU-Net V2 architecture, a well-regarded framework in medical image segmentation, and is further enhanced with a custom-designed interactive training module. This allows the system to iteratively refine tumor segmentations based on direct clinician input, ensuring the final contours align with expert judgment. The tool processes contrast-enhanced T1-weighted MRI scans, a standard imaging modality for meningiomas.
In a comprehensive evaluation using 500 MRI scans from the BraTS 2025 Meningioma RT Segmentation Challenge, Interactive-MEN-RT demonstrated substantial improvements compared to existing segmentation methods. The tool achieved impressive Dice similarity coefficients of up to 77.6% and Intersection over Union scores of 64.8%. These results highlight its superior or comparable performance across all tested interaction types, including points, bounding boxes, lassos, and scribbles, when compared to leading interactive segmentation models like SAM-Med3D, MedSAM2, and nnInteractive.
Also Read:
- AI and Advanced Imaging Revolutionize Radiation Oncology for Precision Cancer Treatment
- Advancing Breast Cancer Diagnosis: A Transformer-Based Approach
The study underscores the critical need for clinically tailored segmentation solutions in high-stakes applications such as meningioma radiotherapy planning. By providing accurate segmentation with minimal user input, Interactive-MEN-RT has the potential to significantly reduce inter-observer variability, streamline RT workflows, and ultimately improve patient outcomes. The researchers have made the code publicly available, fostering further development and adoption of this specialized tool. You can find more details about this research in the full paper available here.
