Computer Vision News - July 2021
2 Summary AI Rese ch 6 Deep learning model Image slices from the T1-weighted MRI scans were used to train the model. Data were partitioned in sets with four-fold cross validation, categorized in sets for test, training, validation. The final model evaluation was done on the 20 subjects which means 500 image slices. H eart of the architecture is a 2D U-Net trained on mini-batches by optimizing the binary cross entropy loss using the Adam optimizer. To allow for quicker convergence, batch normalization was used. The model is visualized in Figure 2 below. A differentiable generalization of the dice score was used to evaluate network exploring the similarity between two binary segmentation volumes. How does it perform? In Figure 3 you can see 3D rendering of manual segmentation trials from both readers. What one could observe is the degree of fragmentation as well as how the location of the lesions differs. Readers would disagree on those locations which affected the dice score. Figure 2: The architecture of the 2D network
Made with FlippingBook
RkJQdWJsaXNoZXIy NTc3NzU=