Bel omurgasi mrg teşhisinde derin öğrenme topluluklari

Background: Segmentation of the lumbar spine anatomy can be a challenging task that entails a considerable amount of time and specialized skills, and involves the critical process of classifying the segmented portions of the lumbar region to determine if the disc is normal or abnormal. Method: We proposed a CNN ensemble method that comprises of five distinct variants of UNet model, namely Unet, ResUnet, SegUnet, VGG16Unet, and VGG19Unet, and then compare the method with TransUNet for segmentation. The method employs the image overlay technique which maps the extracted masks onto the corresponding base images. Subsequently, the masks are classified to determine the normalcy or abnormality of the disc using the pretrained ResNet101 architecture as the backbone for the ensemble models, and compare it with pretrained Vision Transformer. Then, compared our CNN-Transformer ensemble with the CNN ensemble methods. Two datasets were used, however, the main dataset comprised of 1525 slices from 515 patients. Results: The ensemble method proposed herein yielded an Intersection over Union (IoU) of 0.9988 and a Dice Coefficient of 0.9870, outperforming TransUNet in the segmentation task. However, Vision transformer achieved greater results in classification with F1 Score of 0.9867 and 0.9578 for abnormal and normal predictions respectively. The image overlay technique remarkably improved the model's predictive accuracy, from 80% to 92%, compared to Vision Transformer's 96%. Conclusion: We have presented an ensemble method that integrates UNet variants into a single approach, which we compared with transformers. A CNN-Transformer ensemble performs better than a CNN ensemble.