Collaborative Deep Learning for Privacy Preserving Diabetic Retinopathy Detection.

Convolutional Neural Networks (CNNs) are an emerging research area for detection of Diabetic Retinopathy (DR) development in fundus images with highly reliable results. However, its accuracy depends on the availability of big datasets to train such a deep network. Due to the privacy concerns, the strict rules on medical data limit accessibility of images in publicly available datasets. In this paper, we propose a collaborative learning approach to train CNN models with multiple datasets while preserving the privacy of datasets in a distributed learning environment without sharing them. First, CNN networks are trained with private datasets, and tested with the same publicly available images. Based on their initial accuracies, the CNN model with the lowest performance among datasets is forwarded to second lowest performed dataset to retrain it using the transfer learning approach. Then, the retrained network is forwarded to next dataset. This procedure is repeated for each dataset from the lowest performed dataset to the highest. With this ascending chain order fashion, the network is retrained again and again using different datasets and its performance is improved over the time. Based on our experimental results on five different retina image datasets, DR detection accuracy is increased to 93.5% compared with the accuracies of merged datasets (84%) and individual datasets (73%, 78%, 83%, 85%).

Protein Crystallization Segmentation and Classification Using Subordinate Color Channel in Fluorescence Microscopy Images.

The accuracy of detecting protein crystals for fluorescence microscopy images is very critical for high throughput and automated systems. Although the trace fluorescent labeling method could highlight protein crystals, reflection and emission from the fluorescence dye is not always due to crystal regions. Therefore, the analysis of the peak wavelength in the emission spectra of a fluorophore may not always yield effective results. In this paper, we show that using the subordinate color intensity corresponding to longer wavelengths than the peak wavelength of the emission spectra could improve the accuracy of protein crystal detection. Hence, we have built a segmentation method based on the percentile intensity of the subordinate color for trace fluorescently labeled (TFL'd) protein crystallization trial images. Compared to using the dominant color channel, our segmentation method on subordinate color channel was able to reduce the misclassification rate of likely-leads or crystals as non-crystals by the percentage of from 9.71% to 2.02% depending on the classifier. Similarly, the accuracy of classifiers were increased by the percentage of from 1.77% to 5.53%. Our method reached around 94% accuracy while keeping misclassification of likely-leads and crystals as non-crystals below 1%. Moreover, to evaluate the generalizability of our method, we have conducted new wet lab experiments on two proteins, Concanavalin A (Con A) and Ab inorganic pyrophosphate (AbIPPase), and the misclassification rate was below 1%. Our experiments show that using the subordinate channel may be more helpful for TFL'd protein trial image classification.

Schema Matching and Data Integration with Consistent Naming on Protein Crystallization Screens.

The data representation as well as naming conventions used in commercial screen files by different companies make the automated analysis of crystallization experiments difficult and time-consuming. In order to reduce the human effort required to deal with this problem, we present an approach for computationally matching elements of two schemas using linguistic schema matching methods and then transform the input screen format to another format with naming defined by the user. This approach is tested on a number of commercial screens from different companies and the results of the experiments showed an overall accuracy of 97 percent on schema matching which is significantly better than the other two matchers we tested. Our tool enables mapping a screen file in one format to another format preferred by the expert using their preferred chemical names.

Tropical Cyclone Intensity Estimation Using a Deep Convolutional Neural Network.

Tropical cyclone intensity estimation is a challenging task as it required domain knowledge while extracting features, significant pre-processing, various sets of parameters obtained from satellites, and human intervention for analysis. The inconsistency of results, significant pre-processing of data, complexity of the problem domain, and problems on generalizability are some of the issues related to intensity estimation. In this study, we design a deep convolutional neural network architecture for categorizing hurricanes based on intensity using graphics processing unit. Our model has achieved better accuracy and lower root-mean-square error by just using satellite images than 'state-of-the-art' techniques. Visualizations of learned features at various layers and their deconvolutions are also presented for understanding the learning process.

Feature analysis for classification of trace fluorescent labeled protein crystallization images.

BACKGROUND: Large number of features are extracted from protein crystallization trial images to improve the accuracy of classifiers for predicting the presence of crystals or phases of the crystallization process. The excessive number of features and computationally intensive image processing methods to extract these features make utilization of automated classification tools on stand-alone computing systems inconvenient due to the required time to complete the classification tasks. Combinations of image feature sets, feature reduction and classification techniques for crystallization images benefiting from trace fluorescence labeling are investigated. RESULTS: Features are categorized into intensity, graph, histogram, texture, shape adaptive, and region features (using binarized images generated by Otsu's, green percentile, and morphological thresholding). The effects of normalization, feature reduction with principle components analysis (PCA), and feature selection using random forest classifier are also analyzed. The time required to extract feature categories is computed and an estimated time of extraction is provided for feature category combinations. We have conducted around 8624 experiments (different combinations of feature categories, binarization methods, feature reduction/selection, normalization, and crystal categories). The best experimental results are obtained using combinations of intensity features, region features using Otsu's thresholding, region features using green percentile G90 thresholding, region features using green percentile G99 thresholding, graph features, and histogram features. Using this feature set combination, 96% accuracy (without misclassifying crystals as non-crystals) was achieved for the first level of classification to determine presence of crystals. Since missing a crystal is not desired, our algorithm is adjusted to achieve a high sensitivity rate. In the second level classification, 74.2% accuracy for (5-class) crystal sub-category classification. Best classification rates were achieved using random forest classifier. CONTRIBUTIONS: The feature extraction and classification could be completed in about 2 s per image on a stand-alone computing system, which is suitable for real time analysis. These results enable research groups to select features according to their hardware setups for real-time analysis.

Super-Thresholding: Supervised Thresholding of Protein Crystal Images.

In general, a single thresholding technique is developed or enhanced to separate foreground objects from background for a domain of images. This idea may not generate satisfactory results for all images in a dataset, since different images may require different types of thresholding methods for proper binarization or segmentation. To overcome this limitation, in this study, we propose a novel approach called "super-thresholding" that utilizes a supervised classifier to decide an appropriate thresholding method for a specific image. This method provides a generic framework that allows selection of the best thresholding method among different thresholding techniques that are beneficial for the problem domain. A classifier model is built using features extracted priori from the original image only or posteriori by analyzing the outputs of thresholding methods and the original image. This model is applied to identify the thresholding method for new images of the domain. We performed our method on protein crystallization images, and then we compared our results with six thresholding techniques. Numerical results are provided using four different correctness measurements. Super-thresholding outperforms the best single thresholding method around 10 percent, and it gives the best performance for protein crystallization dataset in our experiments.

FocusALL: Focal Stacking of Microscopic Images Using Modified Harris Corner Response Measure.

Automated image analysis of microscopic images such as protein crystallization images and cellular images is one of the important research areas. If objects in a scene appear at different depths with respect to the camera's focal point, objects outside the depth of field usually appear blurred. Therefore, scientists capture a collection of images with different depths of field. Focal stacking is a technique of creating a single focused image from a stack of images collected with different depths of field. In this paper, we introduce a novel focal stacking technique, FocusALL, which is based on our modified Harris Corner Response Measure. We also propose enhanced FocusALL for application on images collected under high resolution and varying illumination. FocusALL resolves problems related to the assumption that focus regions have high contrast and high intensity. Especially, FocusALL generates sharper boundaries around protein crystal regions and good in focus images for high resolution images in reasonable time. FocusALL outperforms other methods on protein crystallization images and performs comparably well on other datasets such as retinal epithelial images and simulated datasets.

Optimizing Associative Experimental Design for Protein Crystallization Screening.

The goal of protein crystallization screening is the determination of the main factors of importance to crystallizing the protein under investigation. One of the major issues about determining these factors is that screening is often expanded to many hundreds or thousands of conditions to maximize combinatorial chemical space coverage for maximizing the chances of a successful (crystalline) outcome. In this paper, we propose an experimental design method called "Associative Experimental Design (AED)" and an optimization method includes eliminating prohibited combinations and prioritizing reagents based on AED analysis of results from protein crystallization experiments. AED generates candidate cocktails based on these initial screening results. These results are analyzed to determine those screening factors in chemical space that are most likely to lead to higher scoring outcomes, crystals. We have tested AED on three proteins derived from the hyperthermophile Thermococcus thioreducens, and we applied an optimization method to these proteins. Our AED method generated novel cocktails (count provided in parentheses) leading to crystals for three proteins as follows: Nucleoside diphosphate kinase (4), HAD superfamily hydrolase (2), Nucleoside kinase (1). After getting promising results, we have tested our optimization method on four different proteins. The AED method with optimization yielded 4, 3, and 20 crystalline conditions for holo Human Transferrin, archaeal exosome protein, and Nucleoside diphosphate kinase, respectively.

CrystPro: Spatiotemporal Analysis of Protein Crystallization Images.

Thousands of experiments corresponding to different combinations of conditions are set up to determine the relevant conditions for successful protein crystallization. In recent years, high throughput robotic set-ups have been developed to automate the protein crystallization experiments, and imaging techniques are used to monitor the crystallization progress. Images are collected multiple times during the course of an experiment. Huge number of collected images make manual review of images tedious and discouraging. In this paper, utilizing trace fluorescence labeling, we describe an automated system called CrystPro for monitoring the protein crystal growth in crystallization trial images by analyzing the time sequence images. Given the sets of image sequences, the objective is to develop an efficient and reliable system to detect crystal growth changes such as new crystal formation and increase of crystal size. CrystPro consists of three major steps- identification of crystallization trials proper for spatio-temporal analysis, spatio-temporal analysis of identified trials, and crystal growth analysis. We evaluated the performance of our system on 3 crystallization image datasets (PCP-ILopt-11, PCP-ILopt-12, and PCP-ILopt-13) and compared our results with expert scores. Our results indicate a) 98.3% accuracy and .896 sensitivity on identification of trials for spatio-temporal analysis, b) 77.4% accuracy and .986 sensitivity of identifying crystal pairs with new crystal formation, and c) 85.8% accuracy and 0.667 sensitivity on crystal size increase detection. The results show that our method is reliable and efficient for tracking growth of crystals and determining useful image sequences for further review by the crystallographers.

Autofocusing for Microscopic Images using Harris Corner Response Measure.

One of the difficulties for proper imaging in microscopic image analysis is defocusing. Microscopic images such as cellular images, protein images, etc. need properly focused image for image analysis. A small difference in focal depth affects the details of an object significantly. In this paper, we introduce a novel auto-focusing approach based on Harris Corner Response Measure (HCRM) and compare the performance with some existing auto-focusing methods. We perform our experiments on protein images as well as a simulated image stack to evaluate the performance of our method. Our results show that our HCRM-based technique outperforms other techniques.

Evaluation of Semi-supervised Learning for Classification of Protein Crystallization Imagery.

In this paper, we investigate the performance of two wrapper methods for semi-supervised learning algorithms for classification of protein crystallization images with limited labeled images. Firstly, we evaluate the performance of semi-supervised approach using self-training with naïve Bayesian (NB) and sequential minimum optimization (SMO) as the base classifiers. The confidence values returned by these classifiers are used to select high confident predictions to be used for self-training. Secondly, we analyze the performance of Yet Another Two Stage Idea (YATSI) semi-supervised learning using NB, SMO, multilayer perceptron (MLP), J48 and random forest (RF) classifiers. These results are compared with the basic supervised learning using the same training sets. We perform our experiments on a dataset consisting of 2250 protein crystallization images for different proportions of training and test data. Our results indicate that NB and SMO using both self-training and YATSI semi-supervised approaches improve accuracies with respect to supervised learning. On the other hand, MLP, J48 and RF perform better using basic supervised learning. Overall, random forest classifier yields the best accuracy with supervised learning for our dataset.

Evaluation of Normalization and PCA on the Performance of Classifiers for Protein Crystallization Images.

In this paper, we investigate the performance of classification of protein crystallization images captured during protein crystal growth process. We group protein crystallization images into 3 categories: noncrystals, likely leads (conditions that may yield formation of crystals) and crystals. In this research, we only consider the subcategories of noncrystal and likely leads protein crystallization images separately. We use 5 different classifiers to solve this problem and we applied some data preprocessing methods such as principal component analysis (PCA), min-max (MM) normalization and z-score (ZS) normalization methods to our datasets in order to evaluate their effects on classifiers for the noncrystal and likely leads datasets. We performed our experiments on 1606 noncrystal and 245 likely leads images independently. We had satisfactory results for both datasets. We reached 96.8% accuracy for noncrystal dataset and 94.8% accuracy for likely leads dataset. Our target is to investigate the best classifiers with optimal preprocessing techniques on both noncrystal and likely leads datasets.

Real-Time Protein Crystallization Image Acquisition and Classification System.

In this paper, we describe the design and implementation of a stand-alone real-time system for protein crystallization image acquisition and classification with a goal to assist crystallographers in scoring crystallization trials. In-house assembled fluorescence microscopy system is built for image acquisition. The images are classified into three categories as non-crystals, likely leads, and crystals. Image classification consists of two main steps - image feature extraction and application of classification based on multilayer perceptron (MLP) neural networks. Our feature extraction involves applying multiple thresholding techniques, identifying high intensity regions (blobs), and generating intensity and blob features to obtain a 45-dimensional feature vector per image. To reduce the risk of missing crystals, we introduce a max-class ensemble classifier which applies multiple classifiers and chooses the highest score (or class). We performed our experiments on 2250 images consisting 67% non-crystal, 18% likely leads, and 15% clear crystal images and tested our results using 10-fold cross validation. Our results demonstrate that the method is very efficient (< 3 seconds to process and classify an image) and has comparatively high accuracy. Our system only misses 1.2% of the crystals (classified as non-crystals) most likely due to low illumination or out of focus image capture and has an overall accuracy of 88%.