Deep learning based clinico-radiological model for paediatric brain tumor detection and subtype prediction.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Pediatric Oncology

Background:

• Early diagnosis of pediatric brain tumors is crucial for improving patient outcomes.
• Magnetic Resonance Imaging (MRI) is a key modality for visualizing brain structures and abnormalities.
• Accurate segmentation and classification of pediatric brain tumors are essential for effective treatment planning.

Purpose of the Study:

• To analyze MRI features of pediatric brain tumors.
• To develop an automated segmentation (AS) tool using deep learning for tumor detection and classification.
• To compare the performance of the AS tool with expert radiologist assessments.

Main Methods:

• Utilized a dataset of 94 pediatric brain MRI cases (75 tumors, 19 normal).
• Developed a deep learning algorithm for automated segmentation of pediatric brain tumors.
• Evaluated the model's sensitivity, specificity, PPV, NPV, and accuracy against radiologist findings.
• Assessed segmentation performance using Dice score and Hausdorff95 distance.

Main Results:

• Identified specific MRI features predicting tumor types (e.g., necrosis/hemorrhage for ependymoma).
• Achieved 90% accuracy and 100% specificity in detecting abnormalities.
• Demonstrated substantial agreement between the deep learning model and radiologists (Cohen's kappa = 0.695).
• The model showed high accuracy in predicting MR characteristics and nearly 80% accuracy in tumor type prediction.

Conclusions:

• The developed deep learning model exhibits high accuracy and specificity in predicting MR characteristics of pediatric brain tumors.
• The automated segmentation tool shows potential for assisting radiologists in timely and accurate diagnosis.
• This AI-powered approach can support clinical decision-making, especially for radiologists lacking specialized neuroradiology training.