Virtual Biopsy for Intracranial Tumors Diagnosis on MRI

TL;DR

A Virtual Biopsy framework comprising: MRI-Processor for standardization; Tumor-Localizer employing vision-language models for coarse-to-fine localization via weak supervision; and Adaptive-Diagnoser with a Masked Channel Attention mechanism fusing local discriminative features with global contexts is proposed.

Abstract

Deep intracranial tumors situated in eloquent brain regions controlling vital functions present critical diagnostic challenges. Clinical practice has shifted toward stereotactic biopsy for pathological confirmation before treatment. Yet biopsy carries inherent risks of hemorrhage and neurological deficits and struggles with sampling bias due to tumor spatial heterogeneity, because pathological changes are typically region-selective rather than tumor-wide. Therefore, advancing non-invasive MRI-based pathology prediction is essential for holistic tumor assessment and modern clinical decision-making. The primary challenge lies in data scarcity: low tumor incidence requires long collection cycles, and annotation demands biopsy-verified pathology from neurosurgical experts. Additionally, tiny lesion volumes lacking segmentation masks cause critical features to be overwhelmed by background noise. To address these challenges, we construct the ICT-MRI dataset - the first public biopsy-verified benchmark with 249 cases across four categories. We propose a Virtual Biopsy framework comprising: MRI-Processor for standardization; Tumor-Localizer employing vision-language models for coarse-to-fine localization via weak supervision; and Adaptive-Diagnoser with a Masked Channel Attention mechanism fusing local discriminative features with global contexts. Experiments demonstrate over 90% accuracy, outperforming baselines by more than 20%.

Figures (7)

• Figure 1: (a) Motivation. Traditional stereotactic biopsy for deep intracranial tumors carries inherent risks (hemorrhage, neurological deficits, tumor seeding) and suffers from sampling bias due to spatial heterogeneity. Non-invasive MRI-based pathology prediction is urgently needed for holistic tumor assessment. (b) Challenges. Challenge-A (Acquisition Difficulty): low incidence, long collection cycles, and expert annotation requiring biopsy-verified pathology; Challenge-B (Utilization Difficulty): tiny lesion volumes and lack of segmentation masks cause critical features to be overwhelmed by background noise. (c) Solutions. We construct the ICT-MRI dataset—the first public biopsy-verified benchmark with 249 cases across four categories; and develop a Virtual Biopsy framework with three modules (MRI-Processor, Tumor-Localizer, Adaptive-Diagnoser with MCA) achieving >90% accuracy, outperforming baselines by over 20%.
• Figure 2: Flowchart of the Whlole Module
• Figure 3: The flowchart of Intracranial Tumor Localization Module (Tumor-Localizer). Block 1 (VLM-guided Spatial Coarse-grained Localization) consists of three sequential steps: (i) Perform axial slicing to convert the preprocessed 3D MRI into a sequence of 2D cross-sectional images; (ii) Feed each 2D slice along with structured radiological prompts to the Vision-Language Model (Qwen3-VL) to predict bounding boxes based on abnormal signal patterns and anatomical asymmetry; (iii) Spatially stack all bounding boxes and apply Gaussian smoothing with morphological dilation to generate a smooth 3D spatial prior map, serving as initial coarse-grained attention. Block 2 (Feature-driven Fine-grained Localization) comprises two steps: (i) Utilize the coarse spatial prior as pseudo-labels to train a lightweight voxel-wise MLP classifier on high-confidence regions, learning intrinsic tumor-background discriminative patterns; (ii) Perform union validation between MLP predictions and VLM-based results to maximize recall and produce the final refined spatial prior map that accurately captures irregular tumor boundaries.
• Figure 4: The flowchart of Adaptive Diagnostic Module (Adaptive-Diagnoser). The module comprises three sequential components: (i) Input: Extract intermediate feature maps from the preprocessed whole-brain MRI using a 3D-CNN backbone, while obtaining the refined spatial mask from the Tumor-Localizer; (ii) Masked Channel Attention (MCA): Perform masked average pooling exclusively on tumor regions to compute channel-wise statistics that extract pure tumor semantics without background interference, then generate adaptive channel weights through a MLP and recalibrate the feature maps via channel-wise multiplication to enhance tumor-specific discriminative channels while suppressing background noise; (iii) Global–Tumor Semantic Fusion: Aggregate global context features and refined tumor-specific features through Global Average Pooling, and produce the final probability distribution.
• Figure 5: Brain MRI Tumor Annonation