A Self-Refining Framework for Intracranial Primary Tumors Diagnosis

Area of Science:

• Neuroimaging
• Artificial Intelligence in Medicine
• Medical Image Analysis

Background:

• Accurate preoperative diagnosis of intracranial tumors via MRI is crucial for effective surgical planning and treatment.
• Current MRI-based diagnostic methods face challenges with inherent class imbalance and image resolution variations.

Purpose of the Study:

• To develop a novel self-refining framework to overcome limitations in MRI-based intracranial tumor diagnosis.
• To improve the accuracy and robustness of tumor segmentation and classification from multi-modal MRI data.

Main Methods:

• Proposed a framework integrating panoptic segmentation for unified tumor and anatomy representation.
• Implemented patch-wise cross-modality attention for adaptive feature fusion from multi-modal MRI.
• Introduced a dynamic loss function to automatically rebalance learning and prioritize rare tumor subtypes.

Main Results:

• Achieved state-of-the-art performance across multiple evaluation metrics for intracranial tumor diagnosis.
• Demonstrated consistent superiority over existing architectures.
• Showcased robustness to varying image resolutions, outperforming current methods.

Conclusions:

• The novel framework enhances detection of diagnostically challenging cases and provides anatomically plausible tumor segmentation.
• The method reduces reliance on uniformly high-resolution scans, offering significant clinical value for real-world deployment.
• This approach improves preoperative diagnosis, aiding surgical planning and therapeutic decision-making for brain tumors.