Challenges such as class imbalance, time intensive visual scoring, and limited amounts of labeled data are often encountered while accessing lung cancer screening low-dose computed tomography (LDCT) data for automated emphysema detection. To tackle these issues, we propose a generative adversarial network (GAN) based on unsupervised anomaly detection (UAD) to identify emphysema in LDCT. And to initiate disease specific feature learning, we introduce data augmentation based on minimum intensity projection (minIP) for the adversarial framework. The generator of the UAD model resembles an image latent-image configuration consisting of fully convolutional encoder-decoder architecture with skip connections. The discriminator is an encoder, which acts as a feature extractor and a classifier. We tested the robustness of the minIP based UAD model by adding anomalies to the training set. Furthermore to illustrate the usefulness of minIP and to serve as a comparison methodology for emphysema detection in LDCT, the proposed model was compared to the 2.5D transfer learning model. Visualization techniques like post processed residual images and Grad-cam maps were used to explain the UAD model inference. Our proposed minIP based UAD model showed an area under the curve of 0.91 0.02 for emphysema detection in LDCT. The UAD model was robust to inclusion of anomalies up to 3%. The augmentation improved the sensitivity of the transfer learning model by 2% however the UAD model outperformed the transfer learning model. The UAD model with minIP is a potential computer-aided tool for early detection of emphysema in lung cancer screening data.

In classical binary statistical pattern recognition optimality in Neyman-Pearson sense, achieved by a (log) likelihood ratio based classifier, is often desirable. A drawback of a Neyman-Pearson optimal classifier is that it requires full knowledge of the (quotient of the) class-conditional probability densities of the input data, which is often unrealistic. The design of neural net classifiers is data driven, meaning that no explicit use is made of the class-conditional probability densities of the input data. In this paper a proof is presented that a neural net can also be trained to approximate a log-likelihood ratio and be used as a Neyman-Pearson optimal, prior-independent classifier. Properties of the approximation of the log-likelihood ratio are discussed. Examples of neural nets trained on synthetic data with known log-likelihood ratios as ground truth illustrate the results.

Along with the deployment of the Face Recognition Systems (FRS), concerns were raised related to the vulnerability of those systems towards various attacks including morphed attacks. The morphed face attack involves two different face images in order to obtain via a morphing process a resulting attack image, which is sufficiently similar to both contributing data subjects. The obtained morphed image can successfully be verified against both subjects visually (by a human expert) and by a commercial FRS. The face morphing attack poses a severe security risk to the e-passport issuance process and to applications like border control, unless such attacks are detected and mitigated. In this work, we propose a new method to reliably detect a morphed face attack using a newly designed denoising framework. To this end, we design and introduce a new deep Multi-scale Context Aggregation Network (MS-CAN) to obtain denoised images, which is subsequently used to determine if an image is morphed or not. Extensive experiments are carried out on three different morphed face image datasets. The Morphing Attack Detection (MAD) performance of the proposed method is also benchmarked against 14 different state-of-the-art techniques using the ISO-IEC 30107-3 evaluation metrics. Based on the obtained quantitative results, the proposed method has indicated the best performance on all three datasets and also on cross-dataset experiments.