Exploring the Efficacy of Spatial Subsampling Methods for 3D Image Projection for 2D Classification

In this paper we purpose several methods for creating 2D projection image from a 3D scan such as from CT or MR scans, and then look at different types of 2.5D projections with a growing window sizes and see its effects on the efficacy of the machine learning model when combined with the selection algorithms. As such we are using Shi-Tomasi Detectors, Sobel Edge Detectors, Sum of Pixel Intensity, Center Select, and Random Selection as selection algorithms from which we select the most interesting index for a given scan, from which depending on the window size, we stack the neighboring slices in the axial plane to the channel dimension creating a 2.5D image from which we could pass into our ResNet-9 model configured for classification and train the model for a given number of epochs, under the same conditions for all of our tests, this ensures that for all of our experiments the main effector is window size and the selection algorithm. From our tests we found that the Center Select on average outperforming the other selection algorithm, this result indicate that medical information is mostly centered around the scan, and when combined with the saving of 60% less parameters, the purposed method shows a new way of processing 3D Scans for processing with a 2D CNN model.