3.4 Experimental Setup 37 sure how far the anomalous samples are from the normal samples. We would like to test our anomaly detection framework in this scope and evaluate whether the generative models are able to understand class related particularities such as shapes, edges, or spatial position; as well as additional hidden features. The raw dataset is provided by the NLST (National Lung Screening Trial), consisting of high resolution chest tomographies. The input data for our models is the result of a nodule detector, and consists of 3D cubes of 32×32×32 mm3 with a voxel size of 1 mm3. Figure 3.4 shows some nodule examples, highlighting the variation in the data and the difficulty for humans to discriminate healthy from anomalous samples. (a) Healthy samples. (b) Anomalous samples. Figure 3.4: Examples of samples in the dataset with their axial, coronal, and sagital perspective, for (a) 3 different healthy nodules and (b) 3 different nodules identified as anomalous (positive for cancer). For our experiments, the input of the models is a 3D cube of 28×28×28 voxels, the result of a data augmentation process that produces sub patches from the original shape of 32×32×32. For visualisation purposes, Figure 3.5 shows the 3D image as a set of 25 slices of 28 ×28 pixels. Table 3.2 shows how we organise the data into training, validation, and test sets. 3.4.2 Architectures and Hyperparameters Our method framework can in principle be implemented with any type of feedforward neural network architecture. Therefore, we consider neural architectures to
RkJQdWJsaXNoZXIy MjY0ODMw