Accurate preoperative MRI diagnosis of intracranial primary tumors is critical for surgical planning and therapeutic decision-making. This study addresses two fundamental limitations of MRI-based diagnosis: the inherent class imbalance and resolution variations. To address these issues, we propose a novel self-refining framework that integrates three key innovations: (1) panoptic segmentation for unified representation of tumors and brain anatomy, (2) patch-wise cross-modality attention enabling adaptive feature fusion from multi-modal MRI, and (3) a dynamic loss function that automatically rebalances learning to prioritize rare tumor subtypes. Our method achieves state-of-the-art performance across multiple evaluation metrics, demonstrating consistent superiority over existing architectures while maintaining robustness to varying image resolutions. These technical advances can translate to clinical benefits, including improved detection of diagnostically challenging cases, anatomically plausible tumor segmentation for surgical planning, and reduced dependence on uniformly high-resolution scans, making the framework particularly valuable for real-world clinical deployment.

Satisfactory progress has been achieved recently in universal segmentation of CT images. Following the success of vision-language methods, there is a growing trend towards utilizing text prompts and contrastive learning to develop universal segmentation models. However, there exists a significant imbalance in information density between 3D images and text prompts. Moreover, the standard fully connected layer segmentation approach faces significant challenges with handling multiple classes and exhibits poor generalizability. To address these challenges, we propose VOxel Interacting with LAnguage method (VOILA) for universal CT image segmentation. Initially, we align voxels and language into a shared representation space and classify voxels based on cosine similarity. Subsequently, we develop the Voxel-Language Interaction framework to mitigate the impact of class imbalance caused by foreground-background discrepancies and variations in target volumes. Furthermore, a Complexity-Aware Sampling method is proposed to focus on region hard to segment, achieved by generating pseudo heatmaps from a trainable Gaussian mixture distribution. Our results indicate the proposed VOILA is capable to achieve improved performance with reduced parameters and computational cost during training. Furthermore, it demonstrates significant generalizability across diverse datasets without additional fine-tuning.