<p>Medical image analysis for diagnosing and classifying various diseases is considered a complex task in deep learning fields. The inherent difficulty arises from the higher dimension involved in the problem, making the analysis of 3D medical images particularly challenging. Therefore, feature extraction has always been crucial and could scientifically affect the classifier results. The focus of our discussion is the influence of image preprocessing on classification results and the application of convolutional long short-term memory (LSTM) networks in the analysis of medical images. In recent years, there have been growing developments of Deep Recurrent Neural Networks (RNN) and LSTM networks for 3D image analysis, specifically in the context of Magnetic Resonance Images (MRI). Originally designed for sequential data processing, these networks have shown to be effective in accurately processing MR images, given their inherent sequential characteristics. In this study, a preprocessing pipeline containing normalization, cropping, 3D image slice selection, sigmoid correction, and gamma correction steps are proposed for simplifying the problem dimension. Finally, an LSTM architecture is proposed for brain tumor classification. The Multimodal Brain Tumor Segmentation Challenge (BraTS) 2019 and BraTS 2020 datasets are employed, and the proposed method demonstrates an accuracy of 84.9% and 89.1% in overall survival prediction. Furthermore, the method achieves an accuracy of 95.4% and 96.3% in tumor grade classification, respectively. This study presents an approach to categorizing brain tumor images using bidirectional convolutional LSTMs. Additionally, this research proposes a preprocessing pipeline for the BraTS dataset, incorporating both sigmoid and gamma corrections, showing improvement in classifier results.</p>