Explanation-Driven Deep Learning Model for Prediction of Brain Tumour Status Using MRI Image Data

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Oncology
• Computational Neuroscience

Background:

• Deep learning (DL) models show promise in analyzing magnetic resonance imaging (MRI) for brain tumor prediction.
• Existing DL models often lack sufficient explanation, interpretability, and high accuracy.
• A need exists for reliable and interpretable AI tools in neuro-oncology.

Purpose of the Study:

• To develop an explanation-driven DL model for predicting brain tumor subtypes (meningioma, glioma, pituitary) from MRI.
• To enhance the accuracy and interpretability of DL models in medical image analysis.
• To address challenges posed by low-quality MRI scans, including noise and metal artifacts.

Main Methods:

• Utilized a dual-input convolutional neural network (CNN) architecture.
• Integrated Local Interpretable Model-Agnostic Explanation (LIME) and Shapley Additive Explanation (SHAP) for model interpretability.
• Introduced Gaussian noise to the MRI dataset to simulate and overcome image quality issues.

Main Results:

• Achieved a classification accuracy of 94.64%, surpassing state-of-the-art methods.
• SHAP values provided consistency and local accuracy for interpretation by examining all input combinations.
• LIME offered localized explanations by constructing sparse linear models around predictions.

Conclusions:

• The proposed explanation-driven DL model demonstrates high accuracy and interpretability for brain tumor subtyping.
• The dual-input CNN approach effectively handles noisy and artifact-laden MRI data.
• This method holds significant potential for clinical integration, trust-building, and even mass screening in resource-limited settings.