Hackers and spammers are employing innovative and novel techniques to deceive novice and even knowledgeable
internet users. Image spam is one of such technique where the spammer varies and changes some portion of the image
such that it is indistinguishable from the original image fooling the users. This paper proposes a deep learning based
approach for image spam detection using the convolutional neural networks which uses a dataset with 810 natural
images and 928 spam images for classification achieving an accuracy of 91.7% outperforming the existing image
processing and machine learning techniques.

DeepImageSpam: Deep Learning based Image Spam Detection

While it is clear that working to deconstruct and reframe irrational thoughts can help bring positive perspective shifts to those suffering from mental distress, the specific factors that precipitate a moment of cognitive change are unclear, particularly in peer-to-peer support. Through data from an on- line mental health forum, we study patterns of peer support to shed light on how moments of cognitive change happen in online communities. Consistent with psychological literature, we use markers of language associated with moments of change, and explore how temporal changes in sentiment surrounding specific topics can approximate cognitive restructuring processes. In addition, we find that culture mediates the effect of these features. Based on the above findings, we propose a predictive model that can distinguish which conversation thread or a post is associated with a moment of cognitive change, and discuss the implications of our work for forum design and online peer support


Moment of change: Predicting Cognitive Restructuring in Online Mental Health Forums

Deep Learning for ELVO Stroke Detection

Purpose
Emergent large vessel occlusions (ELVOs) — the most disabling acute ischemic strokes — are primarily diagnosed through CT angiography (CTA), but require trained radiologists for rapid interpretation [1]. Recent deep learning advancements, especially in the field of image classification, show great promise to automate ELVO detection, accelerate downstream care delivery, and improve patient outcomes. This study tested the fidelity of 2D convolutional neural networks (CNNs) to detect ELVOs.

Materials and Methods
The dataset consisted of 1,212 CTAs from Rhode Island Hospital and surrounding hospitals during a 35-month period (July 2015 to May 2018); 638 patients were ELVO negative and 574 were positive in the internal carotid, middle cerebral, and basilar arteries. This project utilized both custom and pre-trained 2D CNNs (Figure 1). Accuracy and area under the ROC curve (AUC) were the primary metrics used to evaluate model performance. 

The thin CTA data were standardized and cropped to an anatomical volume extending from the circle of Willis to the brain apex. Each cropped three-dimensional CTA was then collapsed into a stack of three non-overlapping, axial Maximum Intensity Projection (MIP) images, which displayed only the densest (brightest) pixels along the z-axis [2]. Data were split into training (80%), validation (10%), and test (10%) sets, then fed into pre-trained architectures — including Inception v3, DenseNet 121, ResNet-50, and NASNet — and custom CNNs. Hyperparameters were iteratively adjusted to improve accuracy and sensitivity.

Results
Of the architectures tested, pre-trained models significantly outperformed custom CNNs; ResNet-50 pre-trained on ImageNet data was adapted to the task at hand by replacing the top layers with three fully connected layers, achieving a peak validation accuracy of 86.3% and AUC of 0.917.

Discussion
The high accuracy and AUC values achieved demonstrate the feasibility of using a fully automated deep learning–based ELVO detection system to streamline ELVO diagnosis and treatment. We aim to conduct a prospective study that involves integrating these models into the RIH Emergency Department, and hope to perform further analysis on multiphase CTA data. These results can potentially be integrated into a platform that helps physicians at RIH and other clinics to accelerate and improve ELVO diagnosis and treatment. 

Deep Learning for ELVO Stroke Detection

Contrary to most reinforcement learning research, which emphasizes on training a deep neural network to have its output layer to approximate a certain strategy, this paper proposes a revolutionary and a reversed method of reinforcement learning. We call this “Reversed Neural Network”. In short, after we train a deep neural network according to a strategy-and-environment-to-payoff table well enough, then we use back-propagation algorithm and propagate the error between the actual output and the desired output back to the “input layer” of a deep neural network gradually to perform a task similar to “human deduction”. And we view the final “input layer” as the fittest strategy for a neural network in a natural environment. 

Reversed Neural Network (Abstract) - Automatically Finding the Nash Equilibrium

Prostate cancer has one of the highest patient survival rates for cancer. The primary focus of treatment has moved to how to treat without decreasing the quality-of-life of patients. This project seeks to understand the surprisingly not-well-known connection between the radiation a patient receives and the symptoms a patient experiences. We use deep learning methods on small datasets as well as statistical analysis methods to evaluate organ sensitivity. In the biomedical field, data is expensive and not abundant, which leads to small datasets. Transfer learning is one method for handling small datasets, but with specialized networks we can lose its full advantages. We use image flipping and curvature-based interpolation methods to create more data in order to leverage transfer learning. Using interpolated and augmented data, we can train a convolutional autoencoder network to get near-optimal starting points for the weights in our convolutional neural network for analyzing the relationship between patient-reported quality-of-life and radiation. Furthermore, we use analysis of variance and logistic regression to analyze organ sensitivity to radiation and develop dosage thresholds for each organ region. We identify regions of both the bladder and rectum that are highly correlated with changes in individual symptoms. Finally, we estimate radiation therapy dosage thresholds to determine how high radiation therapy dosage needs to be in order to trigger collateral symptoms. Connecting deep learning methods and organ sensitivity provides a framework to inform patient care in the context of their quality-of-life.

Using Quality-of-Life Scores to Guide Prostate Radiation Therapy Dosing

At the Southern California Gas Company, we need to distinguish between various types of abnormal gas usage. SoCalGas has built predictive models to detect such instances. For example, we need to identify  water leaks from gas leaks and respond appropriately.

SoCalGas has very large datasets collected from almost six million residential smart meters in our region.   The data collected for analyses and modeling includes gas consumption,  temperature,  service operation times, meter location etc.  Via customer pattern analyses  data, SoCalGas attempts to identify the different patterns of water leaks.

However, to build predictive models, requires very large sets of accurately labeled and cleaned data.  There are a practical resource limitations to a manual labeling process, especially with consumption data that can be noisy.   Another modeling challenge is that the meters, which are the data sources, are from different families and in different geographical locations. 

For these reasons,  we propose the use of a Variational Autoencoder (VAE) model as a semi-supervised learning method.  With a certain amount of labeled data, we can train a much larger data set a) without bias from partial data and b) reduce the noise in the larger dataset.  This approach has allowed us to differentiate and predict abnormal consumption patterns due to water leak.

The VAE is known as a generative model. The VAE model is and upgraded architecture of a regular autoencoder by replacing the usual deterministic function Q with a probabilistic function q((z|x)).  A VAE model learns soft ellipsoidal regions in latent space by effectively force filling the gaps where labels are missing.  The missing labels are able to be filled via this method when we apply the VAE model to the data.  Also, the VAE model is encoded to the latent space and decoded to reconstruct the data. In this process, just like autoencoder, any noise is reduced.

In experiments, we compared VAE results with our consumption data analytics pattern recognition results and also with autoencoder results. This model showed efficient prediction of missing labels and prediction of properties of various areas. Depending on different ratio of labeled data and non-labeled data, the performance and accuracy of model were affected. 

MIC Submissions

DeepImageSpam: Deep Learning based Image Spam Detection

Moment of change: Predicting Cognitive Restructuring in Online Mental Health Forums

Deep Learning for ELVO Stroke Detection

Reversed Neural Network (Abstract) - Automatically Finding the Nash Equilibrium

Using Quality-of-Life Scores to Guide Prostate Radiation Therapy Dosing

Hot Water Leak Detection Using Variational Autoencoder Model