Figure 1. a) S chematic of the CAE. The CAE receives an IR spectrum as input and tries to reproduce this input data. The pink convolutional layers have a Kernel size of 10 with 32,64, and 32 filters. An additional Dropout layer (20% dropout rate) was added to prevent overfitting. The green layer describes a fully connected dense layer with 24 neurons. The blue part of the network consists of three upscaling convolutional layers with a kernel size of 3,5 and 4, respectively, and 64,128 and 55 Filters. The total number of parameters accounts for 800.863 parameters. b) The green layers represent the two added dense layers with 10 neurons for the classification tasks. The grey squares represent output neurons. The number of output neurons depends on the number of classes to be distinguished. The trainable parameters account for 261.
This final part of the network will be trained to classify the different subtypes of lymphoma and normal (reactive) control. It uses the pre-trained feature detection of the first part of the Autoencoder. With that, the number of parameters to be fitted is reduced to 253 Parameters.
Labelling training and test data are required to train the classifier part of the neural network. Here, areas of interest in lymphoma tissue (FL = follicular and intrafollicular area; DLBCL, non-GCB and DLBCL, GCB subtype) and rLN were labelled, and the corresponding spectra were extracted. One sample served as a training set, and the other as an evaluation set. The labelling procedure of the training data is depicted in Figure 2 .