Indiana University-Purdue University Indianapolis (IUPUI) / Many complex processes can be viewed as sequential events on a network. In this thesis, we study the interplay between a network and the event sequences on it. We first focus on predicting events on a known network. Examples of such include: modeling retweet cascades, forecasting earthquakes, and tracing the source of a pandemic. In specific, given the network structure, we solve two types of problems - (1) forecasting future events based on the historical events, and (2) identifying the initial event(s) based on some later observations of the dynamics. The inverse problem of inferring the unknown network topology or links, based on the events, is also of great important. Examples along this line include: constructing influence networks among Twitter users from their tweets, soliciting new members to join an event based on their participation history, and recommending positions for job seekers according to their work experience. Following this direction, we study two types of problems - (1) recovering influence networks, and (2) predicting links between a node and a group of nodes, from event sequences.

This Master thesis focuses on generative models for medical patient data for radiation therapy. The objective with the project is to implement and investigate the characteristics of a Variational Autoencoder applied to this diverse and versatile data. The questions this thesis aims to answer are: (i) whether the VAE can capture salient features of medical image data, and (ii) if these features can be used to compare similarity between patients. Furthermore, (iii) if the VAE network can successfully reconstruct its input and lastly (iv) if the VAE can generate artificial data having a reasonable anatomical appearance. The experiments carried out conveyed that the VAE is a promising method for feature extraction, since it appeared to ascertain similarity between patient images. Moreover, the reconstruction of training inputs demonstrated that the method is capable of identifying and preserving anatomical details. Regarding the generative abilities, the artificial samples generally conveyed fairly realistic anatomical structures. Future work could be to investigate the VAEs ability to generalize, with respect to both the amount of data and probabilistic considerations as well as probabilistic assumptions. / Fokuset i denna masteruppsats är generativa modeller för patientdata från strålningsbehandling. Syftet med projektet är att implementera och undersöka egenskaperna som en “Variational Autoencoder” (VAE) har på denna typ av mångsidiga och varierade data. Frågorna som ska besvaras är: (i) kan en VAE fånga särdrag hos medicinsk bild-data, och (ii) kan dessa särdrag användas för att jämföra likhet mellan patienter. Därutöver, (iii) kan VAE-nätverket återskapa sin indata väl och slutligen (iv) kan en VAE skapa artificiell data med ett rimligt anatomiskt utseende. De experiment som utfördes pekade på att en VAE kan vara en lovande metod för att extrahera framtydande drag hos patienter, eftersom metoden verkade utröna likheter mellan olika patienters bilder. Dessutom påvisade återskapningen av träningsdata att metoden är kapabel att identifiera och bevara anatomiska detaljer. Vidare uppvisade generellt den artificiellt genererade datan, en realistisk anatomisk struktur. Framtida arbete kan bestå i att undersöka hur väl en VAE kan generalisera, med avseende på både mängd data som krävs och sannolikhetsteorietiska avgränsningar och antaganden.

Gene-expression profiling enables researchers to quantify transcription levels in cells, thus providing insight into functional mechanisms of diseases and other biological processes. However, because of the high dimensionality of these data and the sensitivity of measuring equipment, expression data often contains unwanted confounding effects that can skew analysis. For example, collecting data in multiple runs causes nontrivial differences in the data (known as batch effects), known covariates that are not of interest to the study may have strong effects, and there may be large systemic effects when integrating multiple expression datasets. Additionally, many of these confounding effects represent higher-order interactions that may not be removable using existing techniques that identify linear patterns. We created Confounded to remove these effects from expression data. Confounded is an adversarial variational autoencoder that removes confounding effects while minimizing the amount of change to the input data. We tested the model on artificially constructed data and commonly used gene expression datasets and compared against other common batch adjustment algorithms. We also applied the model to remove cancer-type-specific signal from a pan-cancer expression dataset. Our software is publicly available at https://github.com/jdayton3/Confounded.

Treating cancer with radiotherapy requires precise planning. Several planning pipelines rely on reference dose mimicking, where one tries to find machine parameters best mimicking a given reference dose. Dose mimicking relies on having a function that quantifies dose similarity well, necessitating methods for feature extraction of dose images. In this thesis we investigate ways of extracting features from clinical doseimages, and propose a few proof-of-concept dose mimicking functions using the extracted features. We extend current techniques and lay the foundation for new techniques for feature extraction, using mathematical frameworks developed in entirely different areas. In particular we give an introduction to wavelet theory, which provides signal decomposition techniques suitable for analysing local structure, and propose two different dose mimicking functions using wavelets. Furthermore, we extend ROI-based mimicking functions to use artificial ROIs, and we investigate variational autoencoders and their application to the clinical dose feature extraction problem. We conclude that the proposed functions have the potential to address certain shortcomings of current dose mimicking functions. The four methods all seem to approximately capture some notion of dose similarity. Used in combination with the current framework they have the potential of improving dose mimickingresults. However, the numerical tests supporting this are brief, and more thorough numerical investigations are necessary to properly evaluate the usefulness of the new dose mimicking functions. / Behandling av cancer med strålterapi kräver precis planering. Flera olika planeringsramverk bygger på doshärmning, som innebär att hitta de maskinparametrar som bäst härmar en given referensdos. För doshärmning behövs en funktion som kvantifierar likheten mellan två doser, vilket kräver ett sätt att extrahera utmärkande egenskaper – så kallade features – från dosbilder. I det här examensarbetet undersöker vi olika matematiska metoder för att extrahera features från kliniska dosbilder, och presenterar några olika förslag på prototyper till doshärmningsfunktioner, konstruerade utifrån extraherade features. Vi utvidgar nuvarande tekniker och lägger grunden för nya tekniker genom att använda matematiska ramverk utvecklade för helt andra syften. Speciellt så ger vi en introduktion till wavelet-teori, som ger matematiska verktyg för att analysera lokala beteenden hos signaler, exempelvis bilder. Vi föreslår två olika doshärmningsfunktioner som utnyttjar wavelets, och utvidgar ROI-baseraddoshärmning genom att introducera artificiella ROIar. Vidare så undersökervi så kallade variational autoencoders  och möjligheten att använda dessa för extrahering av features från dosbilder. Vi kommer fram till att de föreslagna funktionerna har potential att åtgärda vissa begränsningar som finns hos de doshärmningsfunktioner som används idag. De fyra metoderna verkar alla approximativt kvantifiera begreppet doslikhet. Användning av dessa nya metoder i kombination med nuvarande ramverk för doshärmning har potential att förbättra resultaten från doshärmning. De numeriska undersökningar som underbygger dessa slutsatser är dock inte särskilt ingående, så mer noggranna numeriska tester krävs för att kunna ge några definitiva svar angående de presenterade doshärmningsfunktionernas användbarhet ipraktiken.

In this thesis, a comparison of three different pre-processing methods for imbalanced classification data, is conducted. Variational Autoencoder, Random Under-Sampling Boosting and a hybrid approach of the two, are applied to three imbalanced classification data sets with different class imbalances. A logistic regression (LR) model is fitted to each pre-processed data set and based on its classification performance, the pre-processing methods are evaluated. All three methods shows indications of different advantages when handling class imbalances. For each pre-processed data, the LR-model has is better at correctly classifying minority class observations, compared to a LR-model fitted to the original class imbalanced data sets. Evaluating the overall classification performance, both VAE and RUSBoost shows improving classification results while the hybrid method performs worse for the moderate class imbalanced data and best for the highly imbalanced data.

Motivated by successes in the image generation domain, this thesis presents a novel Hybrid VQ-VAE (H-VQ-VAE) approach for generating realistic synthetic time series data with categorical features. The primary motivation behind this work is to address the limitations of existing generative models in accurately capturing the underlying structure and patterns of time series data, especially when dealing with categorical features.  Our proposed H-VQ-VAE model builds upon the foundation of the VQ-VAE architecture and consists of two separate VQ-VAEs: the whole VQ-VAE and the sliding VQ-VAE. Both models share a ResNet-based architecture with conv1d layers to effectively capture the temporal structure within the time series data. The whole VQ-VAE focuses on entire sequences of data to learn relationships between categorical and numerical features, while the sliding VQ-VAE exclusively processes numerical features using a sliding window approach. We conducted experiments on multiple datasets to evaluate the performance of our H-VQ-VAE model in comparison with the original VQ-VAE and TimeGAN models. Our evaluation used a train-on-real and test-on-synthetic approach, focusing on metrics such as Mean Absolute Error (MAE) and Explained Variance (EV). The H-VQ-VAE model achieved a 25-50% better MAE for numerical features compared to the VQ-VAE and outperformed TimeGAN by 45-75% on the complex dataset indicating its effectiveness in capturing the underlying structure and patterns of the time series data. In conclusion, the H-VQ-VAE model offers a promising approach for generating realistic synthetic time series data with categorical features, with potential applications in various fields where accurate data generation is crucial.

Anomaly detection is the process of detecting samples in a dataset that are atypical or abnormal. Anomaly detection can for example be of great use in an industrial setting, where faults in the manufactured products need to be detected at an early stage. In this setting, the available image data might be from different non-natural domains, such as the depth domain. However, the amount of data available is often limited in these domains. This thesis aims to investigate if a convolutional neural network (CNN) can be trained to perform anomaly detection well on limited datasets in non-natural image domains. The attempted approach is to train the CNN as an autoencoder, in which the CNN is the encoder network. The encoder is then extracted and used as a feature extractor for the anomaly detection task, which is performed using Semantic Pyramid Anomaly Detection (SPADE). The results are then evaluated and analyzed. Two autoencoder models were used in this approach. As the encoder network, one of the models uses a MobileNetV3-Small network that had been pretrained on ImageNet, while the other uses a more basic network, which is a few layers deep and initialized with random weights. Both these networks were trained as regular convolutional autoencoders, as well as variational autoencoders. The results were compared to a MobileNetV3-Small network that had been pretrained on ImageNet, but had not been trained as an autoencoder. The models were tested on six different datasets, all of which contained images from the depth and intensity domains. Three of these datasets additionally contained images from the scatter domain, and for these datasets, the combination of all three domains was tested as well. The main focus was however on the performance in the depth domain. The results show that there is generally an improvement when training the more complex autoencoder on the depth domain. Furthermore, the basic network generally obtains an equivalent result to the more complex network, suggesting that complexity is not necessarily an advantage for this approach. Looking at the different domains, there is no apparent pattern to which domain yields the best performance. This rather seems to depend on the dataset. Lastly, it was found that training the networks as variational autoencoders did generally not improve the performance in the depth domain compared to the regular autoencoders. In summary, an improved anomaly detection was obtained in the depth domain, but for optimal anomaly detection with regard to domain and network, one must look at the individual datasets. / <p>Examensarbetet är utfört vid Institutionen för teknik och naturvetenskap (ITN) vid Tekniska fakulteten, Linköpings universitet</p>

Style transfer using Generative adversarial networks (GANs) has been successful in recent publications. One field in style transfer is music style transfer, in which a piece of music is transformed in some way, be it through genre-, harmonic-, rhythmic transfer, etc. In this thesis, we have performed genre style transfer using a CycleGAN architecture and symbolic representation of data. Previous work using the same architecture and representation has focused solely on transferring the arrangement of the notes (composition). We have improved this work by including the transfer of multiple instruments (timbre) to create more convincing results. Additional discriminators were added to the CycleGAN architecture to achieve this, and they are individually tasked with enforcing the timbre and composition of a song. Previous works have also used variable autoencoders (VAEs) with sequential data representation for style transfer. The use of VAEs for genre style transfer using symbolic data representation instead of sequential was explored, and recommendations for future work include omitting faults found during exploration. Two different classifiers were created to evaluate the results of the CycleGAN model. One uses symbolic representation, in which all instruments are merged into one, thus evaluating the composition of the generated songs. The other classifier uses a spectrogram representation which evaluates the transfer as a whole, both timbre and composition. The evaluation of the improved CycleGAN model using the classifiers showed that it could perform genre style transfer successfully even when adding timbre to the style transfer.

Here, we show that training with multiple noncommunicable diseases (NCDs) is both feasible and beneficial to modeling this class of diseases. We first use data from the Cancer Genome Atlas (TCGA) to train a pan cancer model, and then characterize the information the model has learned about the cancers. In doing this we show that the model has learned concepts that are relevant to the task of cancer classification. We also test the model on datasets derived independently of the TCGA cohort and show that the model is robust to data outside of its training distribution such as precancerous legions and metastatic samples. We then utilize the cancer model as the basis of a transfer learning study where we retrain it on other, non-cancer NCDs. In doing so we show that NCDs with very differing underlying biology contain extractible information relevant to each other allowing for a broader model of NCDs to be developed with existing datasets. We then test the importance of the samples source tissue in the model and find that the NCD class and tissue source may not be independent in our model. To address this, we use the tissue encodings to create augmented samples. We test how successfully we can use these augmented samples to remove or diminish tissue source importance to NCD class through retraining the model. In doing this we make key observations about the nature of concept importance and its usefulness in future neural network explainability efforts.

Deep learning-based dose prediction is a promising approach to automated radiotherapy planning but carries with it the risk of failing silently when the inputs are highly abnormal compared to the training data. One way to address this issue is to develop a dedicated outlier detector capable of detecting anomalous patient geometries. I examine the potential of so-called generative models to handle this task. These models are promising due to being able to model the distribution of the input data regardless of the downstream task, but they have also been shown to suffer from serious biases when applied to outlier detection. No consensus has been reached regarding the root cause of these biases, or how to address them. I investigate this by attempting to design a variational autoencoder-based outlier detector trained to detect anomalous samples of shapes represented in a binary voxel format. I find the standard procedure application to suffer from severe bias when encountering cropped shapes, leading to systematic misclassification of some outlier patient cases. I overcome this by adopting a segmentation metric as an out-of-distribution metric and show that this outperforms recently proposed general-purpose solutions to the likelihood bias issue. I then benchmark my proposed method on clinical samples and conclude that this approach achieves performance comparable to a one-class support vector machine model that uses handcrafted domain-specific features. / Djupinlärningsbaserad dosprediktion är en mycket lovande metod för att automatiskt generera behandlingsplaner för strålterapi. Djupinlärningsmodeller kan dock endast förväntas fungera på data som är tillräckligt lik träningsdatan, vilket skapar en säkerhetsrisk i kliniska miljöer. Ett möjlig lösning på detta problem är att använda en särskild detektor som klarar av att identifiera avvikande data. I denna uppsats undersöker jag om en generativa djupinlärningsmodell kan användas som en sådan detektor. Generativa modeller är särskilt intressanta för detta ändamål då de är både kraftfulla och flexibla. Dessvärre har generativa modeller visats kunna vilseledas av vissa typer av data. Orsakerna och de underliggande faktorerna till detta har ännu inte identifierats. Jag undersöker denna problematik genom att designa en detektor baserad på en variationell autokodare. Jag upptäcker att den en naiv applikation av denna modell inte är tillräcklig för den kliniska datan, då modellen systematiskt felvärderar beskärda former. Jag löser detta problem genom att nyttja ett modifierat segmenteringsmått som detektionsmått, och visar att denna metod fungerar bättre än mer allmänna lösningar på vilseledningsproblemet. Jag evaluerar metoderna på klinisk data och finner att min metod fungerar lika bra som en en-klass stödvektormaskin som använder sig av handgjorda domänspecifika features.