A Robust CNN-LSTM Framework for Sensitive and Specific Brain Tumor Classifications

Diagnosing brain tumor is a key medical challenge where identification of the tumor is crucial to determine an appropriate treatment methodology. In this paper, we present a new hybrid deep learning model that combines Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks to take advantage of spatial and temporal features on multimodal MRI scans to better diagnose classifying diseases. It combines a CNN component to extract high level spatial features and an LSTM component to encode temporal dependencies between MRI slice features, thereby making a robust diagnostic system. The model was trained and validated using the publicly available BraTS dataset, achieving an accuracy of 97.8<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">%</sup>, sensitivity of 98.5%, and specificity of 97.3%, outperforming state-of-the-art models such as ResNet50 (accuracy: 95. It was compared to a NeXt 3D (accuracy: 96.5%, 91.3% sensitivity, 99.8% specificity) and VGG16 (accuracy: 94.7%, 94.9% sensitivity, 94.3% specificity). An ablation study revealed a contribution of each module: the accuracy of the standalone CNN module is 92.4%, the standalone LSTM module is 90.1%, and the accuracy of the proposed hybrid model is 94.0%. The model's focus on tumor regions was validated by Grad-CAM visualizations that improved its interpretability for clinical applications. Both why and how it performed were proven through statistical analysis including confidence intervals and paired t-tests of p < 0.01. Our resulting CNN-LSTM hybrid model is proven by data as a reliable and interpretable means of brain tumor diagnosis that provides high diagnostic accuracy while being clinically interpretable. The future work is to experiment with integrating attentions and larger data sets to improve performance and scalability.