Deep Vision for Breast Cancer Classification and Segmentation

(1) Background: Female breast cancer diagnoses odds have increased from 11:1 in 1975 to 8:1 today. Mammography false positive rates (FPR) are associated with overdiagnoses and overtreatment, while false negative rates (FNR) increase morbidity and mortality. (2) Methods: Deep vision supervised learning classifies 299 × 299 pixel de-noised mammography images as negative or non-negative using models built on 55,890 pre-processed training images and applied to 15,364 unseen test images. A small image representation from the fitted training model is returned to evaluate the portion of the loss function gradient with respect to the image that maximizes the classification probability. This gradient is then re-mapped back to the original images, highlighting the areas of the original image that are most influential for classification (perhaps masses or boundary areas). (3) Results: initial classification results were 97% accurate, 99% specific, and 83% sensitive. Gradient techniques for unsupervised region of interest mapping identified areas most associated with the classification results clearly on positive mammograms and might be used to support clinician analysis. (4) Conclusions: deep vision techniques hold promise for addressing the overdiagnoses and treatment, underdiagnoses, and automated region of interest identification on mammography.

Pre-stack Bayesian lithofacies classification technology and application in oil-sand reservior prediction

The oil sand reservoirs in the Athabasca region of Canada are estuarine deposits affected by tides. The strata are inclined and the interlayers are well-developed. Accurate spatial characterization of reservoirs and interlayers is the key for efficient oil-sand development. In this paper, we use pre-stack Bayesian lithofacies classification technology to predict the spatial distribution characteristics of reservoirs and interlayers of oil-sand reservoirs. We first use log lithofacies data as a label, select lithofacies sensitive elastic parameters to make a lithofacies classification probability distribution cross-plot, and then project the lithofacies-sensitive elastic parameter volumes into the lithofacies classification probability distribution cross-plot. Finally, we predict the spatial probability distribution of different lithofacies. Probabilistic characterization can enhance the recognition of transitional lithology and thin layers in the inversion results, reduce the uncertainty in the prediction of reservoirs and interlayers, and significantly improve the prediction accuracy of reservoirs and interlayers. The field application results in the Kinosis study area show that the probability volume predicted by this technology can distinguish interlayers greater than 1 meter thick and identify interlayers greater than 2 meters thick, which meets the technical requirements of oil-sand SAGD (Steam Assisted Gravity Drainage) development.

Melanoma Recognition and Lesion Segmentation using Multi-Instance Learning

Melanoma is one of the deadliest forms of skin cancer, but early and accurate identification can significantly improve the survival rate of patients. In this paper, an end-to-end framework based on multi-instance learning is proposed for melanoma recognition and lesion segmentation simultaneously. To make full use from the information of high-resolution images, we take each image block (super-pixel) as an instance in a bag and use multi-instance learning based on a graph convolutional network to recognize melanoma. Moreover, skin lesion segmentation is derived from attention weights and is calibrated by classification probability vectors. As a result, the AUC of our method for melanoma recognition reaches 0.93, which is much higher compared with other related methods. Also, the Jaccard index (JA) of our method for melanoma-related skin lesion segmentation reaches 0.699. In our end-to-end approach, segmentation and recognition are treated as intimately coupled processes, and hence, a high JA is also an indication of the reliability of melanoma recognition. Collectively, these findings confirmed that our method effectively assists melanoma diagnosis.

Validation of the Relationship Between Iris Color and Uveal Melanoma Using Artificial Intelligence With Multiple Paths in a Large Chinese Population

Previous studies have shown that light iris color is a predisposing factor for the development of uveal melanoma (UM) in a population of Caucasian ancestry. However, in all these studies, a remarkably low percentage of patients have brown eyes, so we applied deep learning methods to investigate the correlation between iris color and the prevalence of UM in the Chinese population. All anterior segment photos were automatically segmented with U-NET, and only the iris regions were retained. Then the iris was analyzed with machine learning methods (random forests and convolutional neural networks) to obtain the corresponding iris color spectra (classification probability). We obtained satisfactory segmentation results with high consistency with those from experts. The iris color spectrum is consistent with the raters’ view, but there is no significant correlation with UM incidence.

Systematic Water Fraction Estimation for a Global and Daily Surface Water Time-Series

Fresh water is a vital natural resource. Earth observation time-series are well suited to monitor corresponding surface dynamics. The DLR-DFD Global WaterPack (GWP) provides daily information on globally distributed inland surface water based on MODIS (Moderate Resolution Imaging Spectroradiometer) images at 250 m spatial resolution. Operating on this spatiotemporal level comes with the drawback of moderate spatial resolution; only coarse pixel-based surface water quantification is possible. To enhance the quantitative capabilities of this dataset, we systematically access subpixel information on fractional water coverage. For this, a linear mixture model is employed, using classification probability and pure pixel reference information. Classification probability is derived from relative datapoint (pixel) locations in feature space. Pure water and non-water reference pixels are located by combining spatial and temporal information inherent to the time-series. Subsequently, the model is evaluated for different input sets to determine the optimal configuration for global processing and pixel coverage types. The performance of resulting water fraction estimates is evaluated on the pixel level in 32 regions of interest across the globe, by comparison to higher resolution reference data (Sentinel-2, Landsat 8). Results show that water fraction information is able to improve the product’s performance regarding mixed water/non-water pixels by an average of 11.6% (RMSE). With a Nash-Sutcliffe efficiency of 0.61, the model shows good overall performance. The approach enables the systematic provision of water fraction estimates on a global and daily scale, using only the reflectance and temporal information contained in the input time-series.

IoU Regression with H+L-Sampling for Accurate Detection Confidence

It is a common paradigm in object detection frameworks that the samples in training and testing have consistent distributions for the two main tasks: Classification and bounding box regression. This paradigm is popular in sampling strategy for training an object detector due to its intuition and practicability. For the task of localization quality estimation, there exist two ways of sampling: The same sampling with the main tasks and the uniform sampling by manually augmenting the ground-truth. The first method of sampling is simple but inconsistent for the task of quality estimation. The second method of uniform sampling contains all IoU level distributions but is more complex and difficult for training. In this paper, we propose an H+L-Sampling strategy, selecting the high and low IoU samples simultaneously, to effectively and simply train the branch of quality estimation. This strategy inherits the effectiveness of consistent sampling and reduces the training difficulty of uniform sampling. Finally, we introduce accurate detection confidence, which combines the classification probability and the localization accuracy, as the ranking keyword of NMS. Extensive experiments show the effectiveness of our method in solving the misalignment between classification confidence and localization accuracy and improving the detection performance.

Optimizing Reservoir Computers for Signal Classification

Reservoir computers are a type of recurrent neural network for which the network connections are not changed. To train the reservoir computer, a set of output signals from the network are fit to a training signal by a linear fit. As a result, training of a reservoir computer is fast, and reservoir computers may be built from analog hardware, resulting in high speed and low power consumption. To get the best performance from a reservoir computer, the hyperparameters of the reservoir computer must be optimized. In signal classification problems, parameter optimization may be computationally difficult; it is necessary to compare many realizations of the test signals to get good statistics on the classification probability. In this work, it is shown in both a spiking reservoir computer and a reservoir computer using continuous variables that the optimum classification performance occurs for the hyperparameters that maximize the entropy of the reservoir computer. Optimizing for entropy only requires a single realization of each signal to be classified, making the process much faster to compute.

A Study on Enhancement of Fish Recognition Using Cumulative Mean of YOLO Network in Underwater Video Images

In the underwater environment, in order to preserve rare and endangered objects or to eliminate the exotic invasive species that can destroy the ecosystems, it is essential to classify objects and estimate their number. It is very difficult to classify objects and estimate their number. While YOLO shows excellent performance in object recognition, it recognizes objects by processing the images of each frame independently of each other. By accumulating the object classification results from the past frames to the current frame, we propose a method to accurately classify objects, and count their number in sequential video images. This has a high classification probability of 93.94% and 97.06% in the test videos of Bluegill and Largemouth bass, respectively. The proposed method shows very good classification performance in video images taken of the underwater environment.

Robust Image Classification with Cognitive-Driven Color Priors

Existing image classification methods based on convolutional neural networks usually use a large number of samples to learn classification features hierarchically, causing the problems of over-fitting and error propagation layer by layer. Thus, they are vulnerable to adversarial samples generated by adding imperceptible disturbances to input samples. To address the above issue, we propose a cognitive-driven color prior model to memorize the color attributes of target samples inspired by the characteristics of human memory. At inference stage, color priors are indexed from the memory and fused with features of convolutional neural networks to achieve robust image classification. The proposed color prior model is cognitive-driven and has no training parameters, thus it has strong generalization and can effectively defend against adversarial samples. In addition, our method directly combines the features of the prior model with the classification probability of the convolutional neural network, without changing the network structure and its parameters of the existing algorithm. It can be combined with other adversarial attack defense methods, such as various preprocessing modules such as PixelDefense or adversarial training methods, to improve the robustness of image classification. Experiments on several benchmark datasets show that the proposed method improves the anti-interference ability of image classification algorithms.