An active learning-enabled annotation system for clinical named entity recognition

Background

Active learning (AL) has shown the promising potential to minimize the annotation cost while maximizing the performance in building statistical natural language processing (NLP) models. However, very few studies have investigated AL in a real-life setting in medical domain.

Methods

In this study, we developed the first AL-enabled annotation system for clinical named entity recognition (NER) with a novel AL algorithm. Besides the simulation study to evaluate the novel AL algorithm, we further conducted user studies with two nurses using this system to assess the performance of AL in real world annotation processes for building clinical NER models.

Results

The simulation results show that the novel AL algorithm outperformed traditional AL algorithm and random sampling. However, the user study tells a different story that AL methods did not always perform better than random sampling for different users.

Conclusions

We found that the increased information content of actively selected sentences is strongly offset by the increased time required to annotate them. Moreover, the annotation time was not considered in the querying algorithms. Our future work includes developing better AL algorithms with the estimation of annotation time and evaluating the system with larger number of users.

     

A new MLC segmentation algorithm/software for step-and-shoot IMRT delivery

We present a new MLC segmentation algorithm/software for step-and-shoot IMRT delivery. Our aim in this work is to shorten the treatment time by minimizing the number of segments. Our new segmentation algorithm, called SLS (an abbreviation for static leaf sequencing), is based on graph algorithmic techniques in computer science. It takes advantage of the geometry of intensity maps. In our SLS approach, intensity maps are viewed as three-dimensional (3-D) "mountains" made of unit-sized "cubes." Such a 3-D "mountain" is first partitioned into special-structured submountains using a new mixed partitioning scheme. Then the optimal leaf sequences for each submountain are computed by either a shortest-path algorithm or a maximum-flow algorithm based on graph models. The computations of SLS take only a few minutes. Our comparison studies of SLS with CORVUS (both the 4.0 and 5.0 versions) and with the Xia and Verhey segmentation methods on Elekta Linac systems showed substantial improvements. For instance, for a pancreatic case, SLS used only one-fifth of the number of segments required by CORVUS 4.0 to create the same intensity maps, and the SLS sequences took only 25 min to deliver on an Elekta SL 20 Linac system in contrast to the 72 min for the CORVUS 4.0 sequences (a three-fold improvement). To verify the accuracy of our new leaf sequences, we conducted film and ion-chamber measurements on phantom. The results showed that both the intensity distributions as well as dose distributions of the SLS delivery match well with those of CORVUS delivery. SLS can also be extended to other types of Linac systems.     

A scatter-corrected list-mode reconstruction and a practical scatter/random approximation technique for dynamic PET imaging

We describe an ordinary Poisson list-mode expectation maximization (OP-LMEM) algorithm with a sinogram-based scatter correction method based on the single scatter simulation (SSS) technique and a random correction method based on the variance-reduced delayed-coincidence technique. We also describe a practical approximate scatter and random-estimation approach for dynamic PET studies based on a time-averaged scatter and random estimate followed by scaling according to the global numbers of true coincidences and randoms for each temporal frame. The quantitative accuracy achieved using OP-LMEM was compared to that obtained using the histogram-mode 3D ordinary Poisson ordered subset expectation maximization (3D-OP) algorithm with similar scatter and random correction methods, and they showed excellent agreement. The accuracy of the approximated scatter and random estimates was tested by comparing time activity curves (TACs) as well as the spatial scatter distribution from dynamic non-human primate studies obtained from the conventional (frame-based) approach and those obtained from the approximate approach. An excellent agreement was found, and the time required for the calculation of scatter and random estimates in the dynamic studies became much less dependent on the number of frames (we achieved a nearly four times faster performance on the scatter and random estimates by applying the proposed method). The precision of the scatter fraction was also demonstrated for the conventional and the approximate approach using phantom studies.     

Comparing Regularized and Non-Regularized Nonlinear Dipole Fit Methods: A Study in a Simulated Sulcus Structure

The inverse problem arising from EEG and MEG is largely underdetermined. One strategy to alleviate this problem is the restriction to a limited number of point-like sources, the focal source model. Although the singular value decomposition of the spatio-temporal data gives an estimate of the minimal number of dipoles contributing to the measurement, the exact number is unknown in advance and noise complicates the reconstruction. Classical non-regularized nonlinear dipole fit algorithms do not give an estimate for the correct number because they are not stable with regard to an overestimation of this parameter. Too many sources may only describe noise but can still attain a large magnitude during the inverse procedure and may be indiscernible from the true sources. This paper describes a nonlinear dipole fit reconstruction algorithm with a new regularization approach for the embedded linear problem, automatically controlled by the noise in the data and the condition of the occuring least square problems. The algorithm is stable with regard to source components which nearly lie in the kernel of the projection or lead field operator and it thus gives an estimate of the unknown number parameter. EEG simulation studies in a simulated sulcus structure are carried out for an instantaneous dipole model and spatial resolution in the sulcus and stability of the new method are compared with a classical reconstruction algorithm without regularization.     

Testing and optimization of a semiautomatic prostate boundary segmentation algorithm using virtual operators

Image analysis tasks such as size measurement and landmark-based registration require the user to select control points in an image. The output of such algorithms depends on the choice of control points. Since the choice of points varies from one user to the next, the requirement for user input introduces variability into the output of the algorithm. In order to test and/or optimize such algorithms, it is necessary to assess the multiplicity of outputs generated by the algorithm in response to a large set of inputs; however, the input of data requires substantial time and effort from multiple users. In this paper we describe a method to automate the testing and optimization of algorithms using "virtual operators," which consist of a set of spatial distributions describing how actual users select control points in an image. In order to construct the virtual operator, multiple users must repeatedly select control points in the image on which testing is to be performed. Once virtual operators are generated, control points for initializing the algorithm can be generated from them using a random number generator. Although an initial investment of time is required from the users in order to construct the virtual operator, testing and optimization of the algorithm can be done without further user interaction. We illustrate the construction and use of virtual operators by testing and optimizing our prostate boundary segmentation algorithm. The algorithm requires the user to select four control points on the prostate as input.     

The number of genotypic assignments on a genealogy I. The method and simple examples

This paper demonstrates how the number of possible genotypicassignments consistent with the rules of Mendelian geneticsand with any known phenotypes can be calculated for an arbitrarygenealogy. This is of interest both in the context of the usesof the Metropolis algorithm for pedigree analysis and in itsown right. Bounds on the number of states for certain regularand random genealogies are also obtained, and further resultswill be given in a later paper.     

Using total-variation regularization for intensity modulated radiation therapy inverse planning with field-specific numbers of segments

Currently, there are two types of treatment planning algorithms for intensity modulated radiation therapy (IMRT). The beamlet-based algorithm generates beamlet intensity maps with high complexity, resulting in large numbers of segments in the delivery after a leaf-sequencing algorithm is applied. The segment-based direct aperture optimization (DAO) algorithm includes the physical constraints of the deliverable apertures in the calculation, and achieves a conformal dose distribution using a small number of segments. However, the number of segments is pre-fixed in most of the DAO approaches, and the typical random search scheme in the optimization is computationally intensive. A regularization-based algorithm is proposed to overcome the drawbacks of the DAO method. Instead of smoothing the beamlet intensity maps as in many existing methods, we include a total-variation term in the optimization objective function to reduce the number of signal levels of the beam intensity maps. An aperture rectification algorithm is then applied to generate a significantly reduced number of deliverable apertures. As compared to the DAO algorithm, our method has an efficient form of quadratic optimization, with an additional advantage of optimizing field-specific numbers of segments based on the modulation complexity. The proposed approach is evaluated using two clinical cases. Under the condition that the clinical acceptance criteria of the treatment plan are satisfied, for the prostate patient, the total number of segments for five fields is reduced from 61 using the Eclipse planning system to 35 using the proposed algorithm; for the head and neck patient, the total number of segments for seven fields is reduced from 107 to 28. The head and neck result is also compared to that using an equal number of four segments for each field. The comparison shows that using field-specific numbers of segments achieves a much improved dose distribution.     

Accounting for Genotyping Errors in Tagging SNP Selection

One limitation of the existing tagging SNP selection algorithms is that they assume the reported genotypes are error free. However, genotyping errors are often unavoidable in practice. Many tagging SNP selection methods depend heavily on the estimated haplotype frequencies. Recent studies have demonstrated that even slight genotyping errors can lead to serious consequences with regard to haplotype reconstruction and frequency estimation. Here we present a tagging SNP selection method that allows for genotyping errors. Our method is a modification of the pair-wise r² tagging SNP selection algorithm proposed by Carlson et al. (2004) . We have replaced the standard EM algorithm in Carlson's method with an EM that accounts for genotyping errors, in an attempt to obtain better estimates of the haplotype frequencies and r² measure. Through simulation studies we compared the performance of our modified algorithm with that of the original algorithm. We found that the number of tags selected by both methods increased with increasing genotyping errors, though our method led to smaller increase. The power of haplotype association tests using the selected tags decreased dramatically with increasing genotyping errors. The power of single marker tests also decreased, but the reduction was not as much as the reduction in power of haplotype tests. When restricting the mean number of tags selected by both methods to be similar to the baseline number, Carlson's method and our method led to similar power for the subsequent haplotype and single marker tests. Our results showed that, by accounting for random genotyping errors, our method can select tagging SNPs more efficiently than Carlson's method. The computer program that implements our modified tagging SNP selection algorithm is available at our web site: http://www.personal.psu.edu/tuy104/ .     

Inverse planning incorporating organ motion

Accurate targeting is important in intensity-modulated radiation therapy (IMRT). The positional uncertainties of structures with respect to the external beams arise in part from random organ motion and patient setup errors. While it is important to improve immobilization and reduce the influence of organ motion, the residual effects should be included in the IMRT plan design. Current inverse planning algorithms follow the conventional approach and include uncertainties by assuming population-based margins to the target and sensitive structures. Margin around a structure represents a "hard boundary" and the fact that a structure has a spatial probability distribution has been completely ignored. With increasing understanding of spatial uncertainties of structures and the technical capability of fine-tuning the dose distribution on an individual beamlet level in IMRT, it seems timely and important to fully utilize the information in the planning process. This will reduce the "effective" margins of the structures and facilitate dose escalation. Instead of specifying a "hard margin," we describe an inverse planning algorithm which takes into consideration positional uncertainty in terms of spatial probability distribution. The algorithm was demonstrated by assuming that the random organ motion can be represented by a three-dimensional Gaussian distribution function. Other probability distributions can be dealt with similarly. In particular, the commonly used "hard margin" is a special case of the current approach with a uniform probability distribution within a specified range. The algorithm was applied to plan treatment for a prostate case and a pancreatic case. The results were compared with those obtained by adding a margin to the clinical target volume. Better sparing of the sensitive structures were obtained in both cases using the proposed method for approximately the same target coverage.     

A leaf sequencing algorithm to enlarge treatment field length in IMRT

With MLC-based IMRT, the maximum usable field size is often smaller than the maximum field size for conventional treatments. This is due to the constraints of the overtravel distances of MLC leaves and/or jaws. Using a new leaf sequencing algorithm, the usable IMRT field length (perpendicular to the MLC motion) can be mostly made equal to the full length of the MLC field without violating the upper jaw overtravel limit. For any given intensity pattern, a criterion was proposed to assess whether an intensity pattern can be delivered without violation of the jaw position constraints. If the criterion is met, the new algorithm will consider the jaw position constraints during the segmentation for the step and shoot delivery method. The strategy employed by the algorithm is to connect the intensity elements outside the jaw overtravel limits with those inside the jaw overtravel limits. Several methods were used to establish these connections during segmentation by modifying a previously published algorithm (areal algorithm), including changing the intensity level, alternating the leaf-sequencing direction, or limiting the segment field size. The algorithm was tested with 1000 random intensity patterns with dimensions of 21 x 27 cm2, 800 intensity patterns with higher intensity outside the jaw overtravel limit, and three different types of clinical treatment plans that were undeliverable using a segmentation method from a commercial treatment planning system. The new algorithm achieved a success rate of 100% with these test patterns. For the 1,000 random patterns, the new algorithm yields a similar average number of segments of 36.9 +/- 2.9 in comparison to 36.6 +/- 1.3 when using the areal algorithm. For the 800 patterns with higher intensities outside the jaw overtravel limits, the new algorithm results in an increase of 25% in the average number of segments compared to the areal algorithm. However, the areal algorithm fails to create deliverable segments for 90% of these patterns. Using a single isocenter, the new algorithm provides a solution to extend the usable IMRT field length from 21 to 27 cm for IMRT on a commercial linear accelerator using the step and shoot delivery method.     

An optimal algorithm for configuring delivery options of a one-dimensional intensity-modulated beam

The problem of generating delivery options for one-dimensional intensity-modulated beams (1D IMBs) arises in intensity-modulated radiation therapy. In this paper, we present an algorithm with the optimal running time, based on the 'rightmost-preference' method, for generating all distinct delivery options for an arbitrary 1D IMB. The previously best known method for generating delivery options for a 1D IMB with N left leaf positions and N right leaf positions is a 'brute-force' solution, which first generates all N! possible combinations of the left and right leaf positions and then removes combinations that are not physically allowed delivery options. Compared with the brute-force method, our algorithm has several advantages: (1) our algorithm runs in an optimal time that is linearly proportional to the total number of distinct delivery options that it actually produces. Note that for a 1D IMB with multiple peaks, the total number of distinct delivery options in general tends to be considerably smaller than the worst case N!. (2) Our algorithm can be adapted to generating delivery options subject to additional constraints such as the 'minimum leaf separation' constraint. (3) Our algorithm can also be used to generate random subsets of delivery options; this feature is especially useful when the 1D IMBs in question have too many delivery options for a computer to store and process. The key idea of our method is that we impose an order on how left leaf positions should be paired with right leaf positions. Experiments indicated that our rightmost-preference algorithm runs dramatically faster than the brute-force algorithm. This implies that our algorithm can handle 1D IMBs whose sizes are substantially larger than those handled by the brute-force method. Applications of our algorithm in therapeutic techniques such as intensity-modulated arc therapy and 2D modulations are also discussed.     

An algorithm for the selection of proper group intervals for histograms representing clinical laboratory data.

Although numerous statistical methods are available, proper definitive technics to deal with the determination of normal values or ranges in clinical chemistry are unresolved and ill-defined. Initially, the ability to the analyst to define a mathematical function that can be used to best fit the random distribution of generated laboratory values still depends on proper grouping of the data. Herein is proposed a method for grouping data, in order to select the proper class interval to be used. Arbitrariness is removed and maximum information about the population being tested can be achieved. The concepts of "regrouping" and "sign reversal" are utilized, and the validity of this technic for constructing the "best" frequency histogram is verified by a general computer algorithm.     

Computer-based lung sound simulation

An algorithm for the simulation of normal and pathological lung sounds is developed. The simulation algorithm is implemented on a personal computer as well as on a digital signal processor system in real time. Normal, bronchial and tracheal breathing sounds can be generated, and continuous and discontinuous adventitious lung sounds can be added. The attributes of the individual sound components, such as loudness, frequency, duration or number of occurrences within one breathing cycle, are controlled by the user. The quality of the simulations is evaluated by sending audio tapes to 15 experienced pulmonary physicians for a formal assessment. Each tape contains five simulated lung sounds and five real lung sounds from a commercially available teaching tape, presented in random order. Simulated lung sounds are slightly better rated in terms of realism and signal quality when compared to the recordings from the teaching tape. The differences are, however, not significant. 13 out of the 15 physicians feel that computer-based lung sound simulators would be a useful and desirable teaching tool for auscultation courses.     

Low resource processing algorithms for laser Doppler blood flow imaging

The emergence of full field laser Doppler blood flow imaging systems based on CMOS camera technology means that a large amount of data from each pixel in the image needs to be processed rapidly and system resources need to be used efficiently. Conventional processing algorithms that are utilized in single point or scanning systems are therefore not an ideal solution as they will consume too much system resource. Two processing algorithms that address this problem are described and efficiently implemented in a field programmable gate array. The algorithms are simple enough to use low system resource but effective enough to produce accurate flow measurements. This enables the processing unit to be integrated entirely in an embedded system, such as in an application-specific integrated circuit. The first algorithm uses a short Fourier transformation length (typically 8) but averages the output multiple times (typically 128). The second method utilizes an infinite impulse response filter with a low number of filter coefficients that operates in the time domain and has a frequency-weighted response. The algorithms compare favorably with the reference standard 1024 point fast Fourier transform in terms of both resource usage and accuracy. The number of data words per pixel that need to be stored for the algorithms is 1024 for the reference standard, 8 for the short length Fourier transform algorithm and 5 for the algorithm based on the infinite impulse response filter. Compared to the reference standard the error in the flow calculation is 1.3% for the short length Fourier transform algorithm and 0.7% for the algorithm based on the infinite impulse response filter.     

An evaluation of the application of the genetic algorithm to the problem of ordering genetic loci on human chromosomes using radiation hybrid data

Author to whom all correspondence should be addressed We consider the problem of ordering detectable genetic locialong a chromosome by minimizing the number of obligatory breaksthat can be inferred from radiation hybrid data. The problembears some resemblance to the travelling-salesman problem, forwhich genetic algorithms have been used with considerable success.We find that the results from other studies on closely relatedproblems are not directly transferable, and although we didfind a genetic algorithm that performed well in this applicationit would appear that this algorithm is highly sensitive to anychanges in the problem. Moreover, a very simple stochastic algorithmperformed almost as well as our much more complicated and computer-intensivegenetic algorithm and it did so in a fraction of the time. Whilewe do not dispute that genetic algorithms can work on largecomplicated problems, the various modifications and fine-tuningnecessary for good performance tend to be highly problem specificand they are often only arrived at after an exhaustive explorationof possibilities. Thus, we would dispute any claim that geneticalgorithms are robust in their form and range of applicability.     

Smartphone application for emergency signal detection

Currently, a number of studies focus on the study and design of new healthcare technologies to improve elderly health and quality of life. Taking advantage of the popularity, portability, and inherent technology of smartphones, we present an emergency application for smartphones, designated as knock-to-panic (KTP). This innovative and novel system enables users to simply hit their devices in order to send an alarm signal to an emergency service. This application is a complete and autonomous emergency system, and can provide an economic, reliable, and unobtrusive method for elderly monitoring or safety protection. Moreover, the simple and fast activation of KTP makes it a viable and potentially superior alternative to traditional ambient assisted living emergency calls. Furthermore, KTP can be further extended to the general population as well and not just be limited for elderly persons. The proposed method is a threshold-based algorithm and is designed to require a low battery power consumption. The evaluation of the performance of the algorithm in collected data indicates that both sensitivity and specificity are above 90%.     

Intralaboratory quality control of hematology. Comparison of two systems.

Two systems for quality control have been compared, viz., the whole-blood control preparation method and the algorithm method using the geometric moving average XB and a new estimator Y. The system involving whole-blood controls has the advantage of simplicity of operation, but the economic cost of commercial preparations is often high. The algorith system has the advantage that results of all the test samples are used in the calculation; to some extent, this provides a buffer against random variation. The number of count-outs in a given channel is related to the precision of the channel, which in turn is a function of the number of determinations and calculations required for that result. An error of around 1% is introduced into the result each time a calculation is performed. A successful quality control scheme should contain elements of both control preparation and algorithm methods.     

基于随机投影的快速稀疏表示人体动作识别方法

为有效解决体域网人体行为动作远程识别系统低功耗和快速准确识别的问题,提出一种基于随机投影的快速稀疏表示人体行为动作识别的方法。该方法基于压缩感知随机投影方式压缩数据,获取待测试样本邻近类中较少最近邻训练样本,构建测试样本稀疏表示时的训练样本集,以期达到最优线性重构测试样本;在降低传感器装置功耗和稀疏表示识别算法计算复杂度基础上,捕捉人体行为动作本质特征信息,提高多类别动作识别率。采用国际公开可穿戴传感器动作识别数据库WARD多类别动作数据,验证所提算法的有效性。实验结果表明,当数据压缩率为50%,所提算法能够获得最高平均识别率（92.78%）,比传统稀疏表示分类算法获得的动作识别率提高近5%,并显著降低其相应的运行时间,能准确稀疏表示多类别人体行为动作信号,有效降低稀疏表示分类算法的计算法复杂度和运行时间,明显提高多类别动作识别率,为构建快速稀疏表示动作识别提供一个新的思路和方法。     

基于随机森林算法的阿尔茨海默症医学影像分类

为实现阿尔茨海默症(AD)的医学影像分类,辅助医生对患者的病情进行准确判断,本研究对采集的34名AD患者、35名轻度认知障碍患者和35名正常对照组成员的功能磁共振影像进行特征提取和分类,具体思路包括:首先利用皮尔逊相关系数计算脑区之间的功能连接,然后采用随机森林算法对被试不同脑区之间的功能连接进行重要性度量及特征选择,最后使用支持向量机分类器进行分类,利用十倍交叉验证估算分类准确率。实验结果显示,随机森林算法可以对功能连接特征进行有效分析,同时得到AD发病过程的异常脑区,基于随机森林和SVM建立的分类模型对AD、轻度认知障碍的识别具有较好的效果,分类准确率可达90.68%,相关结论可以为AD的早期临床诊断提供客观参照。 【关键词】阿尔茨海默症;功能磁共振成像;随机森林;特征选择     

Direct aperture optimization: a turnkey solution for step-and-shoot IMRT

IMRT treatment plans for step-and-shoot delivery have traditionally been produced through the optimization of intensity distributions (or maps) for each beam angle. The optimization step is followed by the application of a leaf-sequencing algorithm that translates each intensity map into a set of deliverable aperture shapes. In this article, we introduce an automated planning system in which we bypass the traditional intensity optimization, and instead directly optimize the shapes and the weights of the apertures. We call this approach "direct aperture optimization." This technique allows the user to specify the maximum number of apertures per beam direction, and hence provides significant control over the complexity of the treatment delivery. This is possible because the machine dependent delivery constraints imposed by the MLC are enforced within the aperture optimization algorithm rather than in a separate leaf-sequencing step. The leaf settings and the aperture intensities are optimized simultaneously using a simulated annealing algorithm. We have tested direct aperture optimization on a variety of patient cases using the EGS4/BEAM Monte Carlo package for our dose calculation engine. The results demonstrate that direct aperture optimization can produce highly conformal step-and-shoot treatment plans using only three to five apertures per beam direction. As compared with traditional optimization strategies, our studies demonstrate that direct aperture optimization can result in a significant reduction in both the number of beam segments and the number of monitor units. Direct aperture optimization therefore produces highly efficient treatment deliveries that maintain the full dosimetric benefits of IMRT.