Deep Learning Based Image Classification For Lumbar Spine Disorders Using Resnet50 and Data Augmentation Techniques

The research presents a new approach to image classification using deep learning methods. It aims to improve the performance of the pretrained model ResNet50 in classifying images into the categories of LSD, OSF, Spider, and TSEG. The first stage focuses on improving generalization through the processing and augmentation of image datasets. This data is organized into training, validation, and testing sets using a predetermined directory structure. In the OSF category, the performance of ResNet50 is further enhanced using shift, rotation, and zoom augmentation techniques. To refine the model, the final layers of ResNet50 are replaced, adding fully connected layers and a multi class categorization SoftMax output. The model is trained with an Adam optimizer and a categorical cross entropy loss function, using automatic callbacks including early stopping, learning rate reduction, and model checkpointing to improve training. The model is evaluated through accuracy, loss, and a confusion matrix as key performance indicators on the test set. The results for LSD, OSF, Spider, and TSEG are validated using classification accuracy, specificity, sensitivity, and F Score. Additionally, an accuracy of ninety nine percent was achieved using Support Vector Machine, ninety five percent with a Random Forest classifier, and ninety nine percent with Inception version three. The augmented image data and fine tuned pretraining contributed to improved performance on the classification tasks, highlighting the potential of deep learning approaches.