Deep Learning for Pediatric Posterior Fossa Tumor Detection and Classification: A Multi-Institutional Study.

Area of Science:

• Neuro-oncology
• Artificial Intelligence in Medicine
• Medical Imaging Analysis

Background:

• Posterior fossa tumors are the most common brain tumors in children.
• Accurate detection and classification are crucial for effective diagnosis and treatment.
• MR imaging is a primary tool for evaluating these tumors.

Purpose of the Study:

• To develop and evaluate a deep learning model for detecting and classifying pediatric posterior fossa tumors.
• To compare the model's performance against that of experienced radiologists.

Main Methods:

• A multi-institutional cohort of 617 children with posterior fossa tumors (diffuse midline glioma, medulloblastoma, pilocytic astrocytoma, ependymoma) and 199 controls was used.
• A modified ResNeXt-50-32x4d architecture processed T2-weighted MRIs for tumor detection and classification.
• Model performance was benchmarked against four radiologists.

Main Results:

• The deep learning model achieved an AUROC exceeding 0.99 for tumor detection, comparable to radiologists.
• Tumor classification accuracy was 92% with an F1 score of 0.80, outperforming two of the four radiologists.
• The model demonstrated highest accuracy for diffuse midline glioma of the pons, pilocytic astrocytoma, and medulloblastoma.

Conclusions:

• A multi-institutional deep learning model can effectively detect and classify pediatric posterior fossa tumors.
• This AI-driven approach has the potential to augment radiologic diagnosis and improve accuracy.
• Further validation may lead to integration into clinical workflows for pediatric neuro-oncology.