Hybrid CNN–Transformer–LSTM Architectures for Explainable Brain Tumor MRI Classification

Automated classification of brain tumors from magnetic resonance imaging (MRI) can support radiologists by narrowing down differential diagnoses, but clinical usefulness depends on a model being both accurate and interpretable. In this work we design and empirically compare hybrid deep-learning architectures that combine a convolutional neural network (CNN), a Swin-style Transformer sequence refiner, and a bidirectional long short-term memory (BiLSTM) network for four-class brain tumor classification: Glioma, Meningioma, No Tumor, and Pituitary. A full hybrid model (ResNet-50 → Transformer refiner → BiLSTM) trained for fifteen epochs reaches 94.42% test accuracy, a macro F1-score of 94.44%, and a macro one-vs-rest ROC-AUC of 99.63%. To understand which architectural component drives this performance, three two-component variants — CNN+Swin, CNN+LSTM, and a genuine Swin Transformer encoder paired with a BiLSTM (Swin+LSTM) — were trained under an identical three-epoch budget and evaluated on the same held-out test set. Swin+LSTM was consistently the strongest of the three, reaching 90.25% accuracy and a macro F1-score of 90.28%, ahead of CNN+Swin (84.42%) and CNN+LSTM (81.33%). Grad-CAM heatmaps are provided to visualize which regions of an MRI slice drove a given prediction, and the resulting pipeline is deployed through an interactive Streamlit application. The results suggest that pairing a windowed-attention Transformer encoder with a sequence-aggregating recurrent head is a particularly effective and comparatively lightweight combination for this task.