Development of a Fully Optimized Convolutional Neural Network for Astrocytoma Classification in MRI Using Explainable...

Area of Science:

• Neuro-oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Astrocytoma, a common brain glioma, is graded by the WHO, influencing prognosis and treatment.
• Accurate grading is crucial for patient management, particularly for high-malignancy tumors.
• Current grading methods can be subjective and time-consuming.

Purpose of the Study:

• To develop and optimize a Convolutional Neural Network (CNN) for classifying astrocytoma MRI slices into malignant grades (G2-4).
• To evaluate the CNN's performance against state-of-the-art models and assess its robustness for clinical deployment.

Main Methods:

• A fully optimized CNN was trained on 1284 axial 2D MRI slices (T1-weighted contrast-enhanced and FLAIR) from 69 patients.
• Extensive hyperparameter tuning was performed using the Hyperband method.
• Repeated Hold-Out Validation across four data splits was employed, with SHAP and LIME for model interpretability.

Main Results:

• The proposed CNN achieved a mean classification accuracy of 98.05% with an AUC of 0.997.
• The model outperformed state-of-the-art pre-trained models using transfer learning.
• Validation on unmodified slices yielded 93.34% accuracy, demonstrating robustness.

Conclusions:

• The optimized CNN demonstrates high accuracy and robustness in classifying astrocytoma grades from MRI data.
• The AI model shows significant potential for clinical application in neuro-oncology.
• Explainable AI techniques confirmed the model's decision-making process.