کنفرانس مهندسی برق ایران

صفحه اصلی / سی و چهارمین کنفرانس بین المللی مهندسی برق

Classification of Brain tumor patients by Dynamic connectivity and Hybrid CNN–LSTM Models application on fMRI Data

نویسندگان :

Shokoufeh Akbari¹ Karim Abbasian² Somayeh Makouie³ Maryam Shoaran⁴ Farzam Sheikhzadeh⁵

1- دانشکده مهندسی برق و کامپیوتر،دانشگاه تبریز، تبریز، ایران 2- دانشکده مهندسی برق و کامپیوتر،دانشگاه تبریز، تبریز، ایران 3- دانشکده مهندسی برق و کامپیوتر،دانشگاه تبریز، تبریز، ایران 4- دانشکده مهندسی برق و کامپیوتر،دانشگاه تبریز، تبریز، ایران 5- دانشگاه تبریز

کلمات کلیدی :

Tumor and Lesion detection،Resting state fMRI،Dynamic brain connectivity،Deep transfer learning،Neuroimaging،Artificial Neural networks

چکیده :

Functional magnetic resonance imaging (fMRI) provides complex spatiotemporal information about brain activity, requiring models that can capture both spatial and temporal features. In this study, we implemented a deep learning framework combining convolutional neural networks (CNNs) with long short-term memory (LSTM) layers to classify fMRI sequences. Four CNN backbones, VGG16, ResNet50, MobileNetV2, and EfficientNetB0, were integrated with LSTM units to construct hybrid architectures. To ensure methodological rigor and reduce bias, the models were trained and evaluated using k-fold cross-validation, and performance was assessed using Accuracy, Recall, Specificity, and F1-score, along with confusion matrix analysis. The results demonstrated clear differences among the architectures. VGG16 LSTM achieved perfect classification (Accuracy = 1.000, Recall = 1.000, Specificity = 1.000, F1 score = 1.000), though further validation on larger datasets is essential. ResNet50 LSTM produced nearly identical performance (Accuracy = 0.998, Recall = 1.000, Specificity = 0.996, F1 score = 0.998). EfficientNetB0 LSTM achieved balanced outcomes (Accuracy = 0.967, Recall = 1.000, Specificity = 0.936, F1 score = 0.968), while MobileNetV2 LSTM showed weaker performance (Accuracy = 0.935, Recall = 0.893, Specificity = 0.976, F1 score = 0.931). Confusion matrix analysis confirmed that deeper architectures provided near-perfect balance, EfficientNetB0 favored sensitivity, and MobileNetV2 struggled with recall. CNN-LSTM models offer a powerful framework for fMRI classification. VGG16 LSTM and ResNet50 LSTM are recommended for high-stakes neuroimaging applications requiring maximal accuracy, while EfficientNetB0 LSTM provides a modern alternative with strong sensitivity. MobileNetV2 LSTM, although computationally efficient, is less suitable for tasks where diagnostic precision is critical.