فهرست نشریات علمی دسترسی آزاد ایران

Background: Glioblastoma is the most common malignant primary brain tumor in adults. It has been classified into mutant and wild-type subtypes of isocitrate dehydrogenase (IDH). Patient therapeutic options in Glioblastoma vary across subtypes. Therefore, preoperative classification of IDH mutation status is important for therapeutic decision-making. This study aimed to develop and evaluate a metric-meta learning approach based on a Siamese Neural Network model for noninvasive classification of IDH mutation status in patients from magnetic resonance imaging (MRI) scans, based on their similarity. Methods: MRI scans from 93 patients (2018-2024) with glioblastoma tumors were collected and subdivided into IDH-mutant and IDH-wildtype scans. After preprocessing MRI scans, a Siamese Convolutional Neural Network was combined with two pre-trained models, InceptionResNetV2 and ResNet152V2, as an embedding function to differentiate between classes using a small dataset. The network extracts relevant features to classify patients' IDH-mutation status. We propose a self-attention mechanism combined with a Siamese neural network to improve model accuracy. Results: We evaluate a Siamese Neural Network on MRI scans of brain tumors. After adding an attention mechanism to the Siamese Neural network, we obtain classification accuracy (73.33% - 70%), precision (68.42% - 68%), recall (86%- 73.33%), and F1-score (76.20%- 70.96%) on our dataset. Conclusion: Our proposed model demonstrated promising results after combined with an attention mechanism to perform classification of IDH-mutation status using a limited number of brain tumor images and accuracy of this model has improved.