Revolutionizing Brain Tumor Diagnosis: An insight into Convolutional Neural Network Model's Efficacy in Central Nervous System Tumor Analysis with Squash Smear Cytology
WFNS Journal · 2025-01-01 · 02 (01) · e21-e28
DOI: 10.1055/a-2552-6048 ↗Abstract
Abstract Intraoperative squash smear cytology is a valuable diagnostic technique for brain tumors. However, its accuracy and availability are often limited by the need for expert neuropathologists. Artificial Intelligence (AI), particularly Convolutional Neural Networks (CNNs), has the potential to revolutionize this diagnostic process by providing rapid and reliable classifications. This study aims to develop and evaluate a CNN-based AI model for classifying intraoperative squash smear images of gliomas into high-grade and low-grade tumors. A dataset of 10,000 digitally scanned squash smear images (6,000 high-grade and 4,000 low-grade gliomas) was preprocessed and used to train, validate, and test a CNN model. The model's performance was assessed using accuracy, positive predictive value, and F1-score. The CNN model achieved training and validation accuracies of 96.2 and 96.39%, respectively. On the test dataset, it obtained 91% accuracy for high-grade gliomas and 77% for low-grade gliomas. The model also demonstrated a positive predictive value of 86.6% and an F1-score of 0.887. Additionally, feature visualization techniques provided insights into the model's decision-making process. The results indicate that deep learning can serve as a reliable intraoperative diagnostic tool for gliomas, offering neurosurgeons rapid decision support. This approach could significantly improve diagnostic accessibility in centers lacking neuropathological expertise, potentially reducing intraoperative delays and enhancing patient care.
EUVIMED