Multi-Scale CNN-LSTM with Atrous Spatial Pyramid Pooling for Enhanced Brain Tumor Classification in ADC-MRI Sequences

Abstract

Discriminating between malignant and benign brain tumors using ADC-MRI sequences remains challenging due to high similarity in morphology, intra-class heterogeneity, and the sequential nature of the neuroimaging data. This paper introduces a Multi-Scale CNN-LSTM network enhanced by an Atrous Spatial Pyramid Pooling (ASPP) mechanism, bi-directional temporal processing, and squeeze-and-excitation channel attention, which, to the best of our knowledge, is the first model combining multi-scale contextual information with temporal dependency modeling for brain tumor classification using ADC-MRIs. Specifically, local-to-global feature extraction is performed using parallel Inception-based convolutional branches, ASPP is used with variable dilation rates to expand the receptive field without reducing spatial resolution, while Bi-LSTM learns both forward and backward sequential dependencies over seven-slice subsequences of ADC MRIs. A focal loss function was used during training to overcome the issue of imbalanced classes. The proposed model was evaluated using an ADC-MRI dataset of 1,202 images from 41 patients. Results show 98.05% accuracy, 100.00% sensitivity, 95.52% specificity, 96.67% precision, F1-score of 98.31%, AUC-ROC of 0.9993, and MCC of 0.9609. Seven variants ablation experiments prove that ASPP integration provides the largest single-component improvement of 5.89 percentage points (pp) in accuracy compared to the baseline CNN-LSTM, while focal loss gives zero false negative predictions.