Intracranial Hemorrhage (ICH) has high rates of mortality, and risk factors associated with it are sometimes nearly impossible to avoid. Previous techniques to detect ICH using machine learning have shown some promise. However, due to a limited number of labeled medical images available, which often causes poor model accuracy in terms of the Dice coefficient, there is much to be improved. In this paper, we propose a modified u-net and curriculum learning strategy using a multi-task semi-supervised attention-based model, initially introduced by Chen et al., to segment ICH sub-groups from CT images. Using a modified inverse-sigmoid-based curriculum learning training strategy, we were able to stabilize Chen’s algorithm experimentally. This semi-supervised model produced higher Dice coefficient values in comparison to a supervised counterpart, regardless of the amount of labeled data used to train the model. Specifically, when training with 80% of the ground truth data, our semi-supervised model produced a Dice coefficient of 0.67, which was higher than 0.61, obtained by a comparable supervised model. This result also surpassed by a greater margin the one obtained by using the out-of-the-box u-net by Hssayeni et al.

The Loop Current (LC) is the dominant circulation system in the Gulf of Mexico. A long‐term prediction of the LC system (LCS) behavior is critical for understanding the Gulf of Mexico oceanography and ecosystem, and for mitigating outcomes of anthropogenic and natural disasters. In early 2018, the National Academies of Science, Engineering, and Medicine posed a challenge to the research community to develop systems that can forecast the movement of the LCS over longer periods of time than the current state of art. In this paper, a Recurrent Neural Network, the Long Short‐Term Memory (LSTM) network, is applied to predict the LC evolution and the LC ring formation. The LSTM model is trained to learn patterns hidden in sea surface height (SSH) time series. To reduce the memory demand owing to the use of high spatial resolution SSH data set, the region of interest is partitioned into nonoverlapping subregions. After partitioning, an LSTM network is trained to predict the SSH in each subregion. A smoothing function is then applied to reduce discontinuities of the SSH predictions across the partition boundaries, hence error propagation. It is shown that such a machine learning model is capable of predicting the LCS SSH evolution 9 weeks in advance within 40 km in terms of the LCS frontal distance errors. Furthermore, it is shown that the model predicted the timing and general location of eddy Darwin's shedding event 12 weeks in advance, and eddy Cameron's detachment and reattachment 8 weeks in advance.

The traditional process of the complete blood count test has long provided hematologists with information to diagnose blood-based disorders. However, this tedious and manual process can be subject to bias and inaccurate classifications. As a result, automated methods of white-blood cell detection and counting have been sought out as a means to facilitate the process and improve classification accuracy. In this paper, we present a new method, PatternNet-fused Ensemble of Convolutional Neural Networks (PECNN) for classifying white blood cells. The proposed architecture relies on an ensemble method, PatternNet, to fuse outputs of n randomly generated Convolutional Neural Networks. The reliance on randomly generated structures allows the proposed algorithm to be adaptive to data, generalizing its applications. The PatternNet captures the strengths of each participating model while being insensitive to outliers. Through our experimental procedure, we were able to show that our ensemble model outperformed other ensemble models even in midst of noisy data. We were also able to show that the proposed architecture performs as well as a much more sophisticated deep network with a much less computational cost.

Developing effective methods for the automated detection of IDC remains a challenging problem for breast cancer diagnosis. Recently, Cruz and his coworkers proposed a machine learning approach for detection of invasive ductal carcinoma (IDC) from whole image slides containing breast cancer cells. Their method, based on a Convolutional Neural Network (CNN), does not need to handcraft features from images. Their work has the potential of revolutionizing cancer detection, promoting further research and development in this exciting direction. Inspired by Cruz's work, our team investigated various CNN architectures for automated detection of breast cancer. We first implemented a baseline CNN similar to Cruz'ss, and then extended it to four different architectures. All architectures were trained over a large dataset of approximately 275,000, 50x50 RGB image patches. Measurement of quantitative results were done through a global average of 10-Fold Cross Validation tests. Each methodology used the performance measures, F-Measure (F1), and Balanced Accuracy (BAC). One of the fine-tuned CNN architectures yielded the best results in F1, BAC and accuracy (92%, 87%, and 89%, respectively). We also discovered through this study that data augmentation was not effective in automatic detection of breast cancers with the given dataset.

A divide-and-conquer (DAC) machine learning approach was first proposed by Wang et al. to forecast the sea surface height (SSH) of the Loop Current System (LCS) in the Gulf of Mexico. In this DAC approach, the forecast domain was divided into non-overlapping partitions, each of which had their own prediction model. The full domain SSH prediction was recovered by interpolating the SSH across each partition boundaries. Although the original DAC model was able to predict the LCS evolution and eddy shedding more than two months and three months in advance, respectively, growing errors at the partition boundaries negatively affected the model forecasting skills. In the study herein, a new partitioning method, which consists of overlapping partitions is presented. The region of interest is divided into 50%-overlapping partitions. At each prediction step, the SSH value at each point is computed from overlapping partitions, which significantly reduces the occurrence of unrealistic SSH features at partition boundaries. This new approach led to a significant improvement of the overall model performance both in terms of features prediction such as the location of the LC eddy SSH contours but also in terms of event prediction, such as the LC ring separation. We observed an approximate 12% decrease in error over a 10-week prediction, and also show that this method can approximate the location and shedding of eddy Cameron better than the original DAC method.

Simulating the evolution of a coagulating aerosol or cloud of droplets in a key problem in atmospheric science. We present a proof of concept for modeling coagulation processes using a novel combinatorial neural network (CombNN) architecture. Using two types of data from a high‐detail particle‐resolved aerosol simulation, we show that CombNN models outperform standard neural networks and are competitive in accuracy with traditional state‐of‐the‐art sectional models. These CombNN models could have application in learning coarse‐grained coagulation models for multi‐species aerosols and for learning coagulation models from observed size‐distribution data.