<![CDATA[Background: Bone marrow changes in lumbar vertebrae (BMCLVB) are critically linked to lower spinal pain and various spinal disorders, forming a key diagnostic basis. Accurate automated classification using MRI remains a developing area. Objective: This study aimed to develop a hybrid CNN-LSTM model for the automatic classification of BMCLVB from multi-sequence MRI scans. Methods: A newly collected annotated dataset consisting of magnetic resonance imaging scans from 134 patients was utilized, incorporating sagittal T1-weighted, T2-weighted, and STIR sequences. The proposed model uses CNN layers to extract the spatial features from each MRI slice and LSTM layers to capture sequential dependencies across consecutive slices. The dataset was divided into training, validation, and testing groups in a ratio of 70%:20%:10%. The experimental results show that the CNN-LSTM model achieves a classification accuracy of 90.29%, with precision, recall, and F1-score all reaching 91.07%. The proposed approach surpasses various state-of-the-art deep learning models, including CNN, VGG16, VGG19, Xception, ResNet variants, DenseNet201, InceptionResNetV2, NASNetMobile, MobileNetV2, and NASNetLarge. Results: The proposed CNN-LSTM model obtained a test classification accuracy of 90.29%. Precision, recall, and F1-score all reached 91.07%. This performance surpassed that of all benchmark models. Conclusion: The outcomes confirm that the hybrid CNN-LSTM model is highly effective for BMCLVB classification and has strong potential as a clinical decision support tool. This study successfully met its objective, showing the advantage of merged spatial and sequential feature learning for this task. Future work should emphasize extending the dataset and further optimizing the model to improve generalizability across diverse populations and imaging protocols. Keywords: Bone Marrow Changes Lumbar Vertebrae, MRI, Classification, Hybrid CNN-LSTM Model, Medical Image Analysis]]>
