Integrating deep learning and thermal estimation for enhanced MRI-based brain tumor diagnosis

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Accurate brain tumor diagnosis via MRI faces challenges due to reliance on grayscale anatomical data.
• Physiological indicators like tissue temperature, reflecting metabolic activity, are often overlooked in traditional MRI analysis.
• Integrating multimodal data offers potential to enhance diagnostic precision.

Purpose of the Study:

• To develop an AI-driven MATLAB framework integrating MRI, thermal, and textural biomarkers for enhanced brain tumor diagnosis.
• To improve the accuracy of differentiating tumor types (glioma, meningioma, pituitary) and predicting malignancy.
• To explore the correlation between tumor characteristics and estimated tissue temperature.

Main Methods:

• Developed a MATLAB pipeline incorporating deep learning segmentation, morphological analysis, thermal estimation, and texture quantification.
• Utilized a three-layer convolutional neural network (CNN) for classifying tumor types and healthy tissue.
• Derived a formula for temperature estimation based on tumor area: T = 37.0 + 0.7 * log(1 + Area).

Main Results:

• Gliomas exhibited the largest areas and most irregular shapes.
• The CNN model achieved high accuracy (99.2% F1 score) for healthy cases but showed lower recall for pituitary tumors (29%) and precision for meningiomas (48.6%).
• Malignant lesions were estimated to reach temperatures up to 42.2°C, while benign tumors remained below 38.5°C.

Conclusions:

• The integrated AI framework demonstrates potential for improving brain tumor diagnosis by combining multimodal MRI features.
• Further development, including GPU-accelerated training, is needed to refine temperature estimation and address model limitations.
• This multimodal approach holds significant promise for advancing clinical applications and improving patient outcomes.