Neuro-XAI: Explainable deep learning framework based on deeplabV3+ and bayesian optimization for segmentation and...

Area of Science:

• Medical Imaging and Artificial Intelligence
• Computational Neuroscience and Machine Learning

Background:

• Brain tumor diagnosis relies heavily on radiography, but manual image analysis is time-consuming and strenuous for radiologists.
• Machine learning, particularly Convolutional Neural Networks (CNNs), shows promise in medical imaging but faces challenges with hyperparameter tuning, interpretability, and uncertainty quantification.

Purpose of the Study:

• To develop and validate a novel, three-phase deep learning strategy for accurate and interpretable brain tumor segmentation and classification.
• To address the limitations of manual hyperparameter tuning, the 'black box' nature of CNNs, and the lack of uncertainty evaluation in AI-based medical diagnosis.

Main Methods:

• Brain tumor segmentation using the DeeplabV3+ model with Bayesian optimization for hyperparameter tuning.
• Feature extraction from Darknet53 and mobilenetv2 models, followed by Support Vector Machine (SVM) classification with Bayesian-optimized hyperparameters.
• Integration of Explainable Artificial Intelligence (XAI) for model interpretability and confusion entropy for quantifying prediction uncertainty.

Main Results:

• The proposed Bayesian-optimized deep learning framework achieved a 97% classification accuracy and a 0.98 global accuracy.
• The method successfully segmented brain tumors and provided interpretable insights into the CNN's decision-making process.
• Quantification of uncertainty using confusion entropy was successfully implemented, enhancing the reliability of AI-driven diagnoses.

Conclusions:

• The novel three-phase strategy offers a robust and trustworthy AI-based approach for real-time brain tumor diagnosis.
• Combining Bayesian optimization, XAI, and uncertainty quantification significantly improves the performance and clinical applicability of deep learning models in medical imaging.