Improving Brain Tumor Detection by Cortical Surface and Vessels Segmentation Through RGB-to-HSI Transfer Learning.

Area of Science:

• Medical Imaging
• Computational Biology
• Neuroscience

Background:

• Accurate in vivo brain tumor detection via hyperspectral imaging (HSI) is challenging due to tissue complexity and vascularization.
• Conventional methods misclassify tumor tissue as blood vessels, exacerbated by limited annotated data.

Purpose of the Study:

• To develop an improved HSI-based method for accurate in vivo brain tumor detection.
• To overcome limitations of conventional neural network approaches in differentiating tumor tissue from blood vessels.

Main Methods:

• Decomposing the problem into brain cortical surface/blood vessel segmentation and intra-craniotomy tissue segmentation.
• Utilizing pseudo-labels from RGB and HSI data for multimodal training, followed by weakly supervised fine-tuning on HSI data.
• Reframing HSI tissue segmentation as binary (healthy vs. other) within non-overlapped cortex regions.

Main Results:

• Achieved up to a 15.48% increase in F1 score for the tumor class.
• Segmented the brain cortex with a mean Dice Similarity Coefficient (DSC) of 92.08%.
• Accurately detected 95.42% of labeled blood vessel samples in the HSI dataset.

Conclusions:

• The proposed method enhances in vivo brain tumor detection accuracy using HSI.
• The two-stage segmentation approach effectively addresses challenges posed by vascularization and data scarcity.
• This technique offers a more reliable tool for neurosurgical oncology, improving tumor identification and surgical planning.