Comparing ensemble methods combined with different aggregating models using micrograph cell segmentation as an initial application example.

Strategies such as ensemble learning and averaging techniques try to reduce the variance of single deep neural networks. The focus of this study is on ensemble averaging techniques, fusing the results of differently initialized and trained networks. Thereby, using micrograph cell segmentation as an application example, various ensembles have been initialized and formed during network training, whereby the following methods have been applied: (a) random seeds, (b) L 1-norm pruning, (c) variable numbers of training examples, and (d) a combination of the latter 2 items. Furthermore, different averaging methods are in common use and were evaluated in this study. As averaging methods, the mean, the median, and the location parameter of an alpha-stable distribution, fit to the histograms of class membership probabilities (CMPs), as well as a majority vote of the members of an ensemble were considered. The performance of these methods is demonstrated and evaluated on a micrograph cell segmentation use case, employing a common state-of-the art deep convolutional neural network (DCNN) architecture exploiting the principle of the common VGG-architecture. The study demonstrates that for this data set, the choice of the ensemble averaging method only has a marginal influence on the evaluation metrics (accuracy and Dice coefficient) used to measure the segmentation performance. Nevertheless, for practical applications, a simple and fast estimate of the mean of the distribution is highly competitive with respect to the most sophisticated representation of the CMP distributions by an alpha-stable distribution, and hence seems the most proper ensemble averaging method to be used for this application.

Number of necessary training examples for Neural Networks with different number of trainable parameters.

In this work, the network complexity should be reduced with a concomitant reduction in the number of necessary training examples. The focus thus was on the dependence of proper evaluation metrics on the number of adjustable parameters of the considered deep neural network. The used data set encompassed Hematoxylin and Eosin (H&E) colored cell images provided by various clinics. We used a deep convolutional neural network to get the relation between a model's complexity, its concomitant set of parameters, and the size of the training sample necessary to achieve a certain classification accuracy. The complexity of the deep neural networks was reduced by pruning a certain amount of filters in the network. As expected, the unpruned neural network showed best performance. The network with the highest number of trainable parameter achieved, within the estimated standard error of the optimized cross-entropy loss, best results up to 30% pruning. Strongly pruned networks are highly viable and the classification accuracy declines quickly with decreasing number of training patterns. However, up to a pruning ratio of 40%, we found a comparable performance of pruned and unpruned deep convolutional neural networks (DCNN) and densely connected convolutional networks (DCCN).

A deep learning approach to radiation dose estimation.

Currently methods for predicting absorbed dose after administering a radiopharmaceutical are rather crude in daily clinical practice. Most importantly, individual tissue density distributions as well as local variations of the concentration of the radiopharmaceutical are commonly neglected. The current study proposes machine learning techniques like Green's function-based empirical mode decomposition and deep learning methods on U-net architectures in conjunction with soft tissue kernel Monte Carlo (MC) simulations to overcome current limitations in precision and reliability of dose estimations for clinical dosimetric applications. We present a hybrid method (DNN-EMD) based on deep neural networks (DNN) in combination with empirical mode decomposition (EMD) techniques. The algorithm receives x-ray computed tomography (CT) tissue density maps and dose maps, estimated according to the MIRD protocol, i.e. employing whole organ S-values and related time-integrated activities (TIAs), and from measured SPECT distributions of 177Lu radionuclei, and learns to predict individual absorbed dose distributions. In a second step, density maps are replaced by their intrinsic modes as deduced from an EMD analysis. The system is trained using individual full MC simulation results as reference. Data from a patient cohort of 26 subjects are reported in this study. The proposed methods were validated employing a leave-one-out cross-validation technique. Deviations of estimated dose from corresponding MC results corroborate a superior performance of the newly proposed hybrid DNN-EMD method compared to its related MIRD DVK dose calculation. Not only are the mean deviations much smaller with the new method, but also the related variances are much reduced. If intrinsic modes of the tissue density maps are input to the algorithm, variances become even further reduced though the mean deviations are less affected. The newly proposed hybrid DNN-EMD method for individualized radiation dose prediction outperforms the MIRD DVK dose calculation method. It is fast enough to be of use in daily clinical practice.

A comparison of methods for adapting [Formula: see text] dose-voxel-kernels to tissue inhomogeneities.

In [Formula: see text] radionuclide therapies, dosimetry is used for determining patient-individual dose burden. Standard approaches provide whole organ doses only. For assessing dose heterogeneity inside organs, voxel-wise dosimetry based on 3D SPECT/CT imaging could be applied. Often, this is achieved by convolving voxel-wise time-activity-curves with appropriate dose-voxel-kernels (DVK). The DVKs are meant to model dose deposition, and can be more accurate if modelled for the specific tissue type under consideration. In literature, DVKs are often not adapted to these inhomogeneities, or simple approximation schemes are applied. For 26 patients, which had previously undergone a [Formula: see text] -PSMA or -DOTATOC therapy, decay maps, mass-density maps as well as tissue-type maps were derived from SPECT/CT acquisitions. These were used for a voxel-based dosimetry based on convolution with DVKs (each of size [Formula: see text]) obtained by four different DVK methods proposed in literature. The simplest only considers a spatially constant soft-tissue DVK (herein named 'constant'), while others either take into account only the local density of the center voxel of the DVK (herein named 'center-voxel') or scale each voxel linearly according to the proper mass density deduced from the CT image (herein named 'density') or considered both the local mass density as well as the direct path between the center voxel and any voxel in its surrounding (herein named 'percentage'). Deviations between resulting dose values and those from full Monte-Carlo simulations (MC simulations) were compared for selected organs and tissue-types. For each DVK method, inter-patient variability was considerable showing both under- and over-estimation of energy dose compared to the MC result for all tissue densities higher than soft tissue. In kidneys and spleen, 'constant' and 'density'-scaled DVKs achieved estimated doses with smallest deviations to the full MC gold standard (∼[Formula: see text] underestimation). For low and high density tissue types such as lung and adipose or bone tissue, alternative DVK methods like 'center-voxel'- and 'percentage'- scaled achieved superior results, respectively. Concerning computational load, dose estimation with the DVK method 'constant' needs about 1.1 s per patient, center-voxel scaling amounts to 1.2 s, density scaling needs 1.4 s while percentage scaling consumes 860.3 s per patient. In this study encompassing a large patient cohort and four different DVK estimation methods, no single DVK-adaption method was consistently better than any other in case of soft tissue kernels. Hence in such cases the simplest DVK method, labeled 'constant', suffices. In case of tumors, often located in tissues of low (lung) or high (bone) density, more sophisticated DVK methods excel. The high inter-patient variability indicates that for evaluating new algorithms, a sufficiently large patient cohort needs to be involved.

Factors affecting accuracy of S values and determination of time-integrated activity in clinical Lu-177 dosimetry.

INTRODUCTION: In any radiotherapy, the absorbed dose needs to be estimated based on two factors, the time-integrated activity of the administered radiopharmaceutical and the patient-specific dose kernel. In this study, we consider the uncertainty with which such absorbed dose estimation can be achieved in a clinical environment. METHODS: To calculate the total error of dose estimation we considered the following aspects: The error resulting from computing the time-integrated activity, the difference between the S-value and the patient specific full Monte Carlo simulation, the error from segmenting the volume-of-interest (kidney) and the intrinsic error of the activimeter. RESULTS: The total relative error in dose estimation can amount to 25.0% and is composed of the error of the time-integrated activity 17.1%, the error of the S-value 16.7%, the segmentation error 5.4% and the activimeter accuracy 5.0%. CONCLUSION: Errors from estimating the time-integrated activity and approximations applied to dose kernel computations contribute about equally and represent the dominant contributions far exceeding the contributions from VOI segmentation and activimeter accuracy.

Interobserver variability of CT angiography for evaluation of aortic annulus dimensions prior to transcatheter aortic valve implantation (TAVI).

OBJECTIVE: Assessment of aortic annulus dimensions prior to transcatheter aortic valve implantation (TAVI) is crucial for accurate prosthesis sizing in order to avoid prosthesis-annulus-mismatch possibly resulting in complications like valve dislodgement, paravalvular regurgitation or annulus rupture. Contrast-enhanced multidetector computed tomography allows 3-dimensional assessment of aortic annulus dimensions. Only limited data exist about its interobserver variability. METHODS: In 100 consecutive patients with symptomatic severe aortic stenosis (51 male, BMI 27±5kg/m(2), age 81±7 years, heart rate 72±15bpm, Logistic Euroscore 31±14%, STS-Score 7±4%), pre-interventional aortic annulus assessment was performed by dual source computed tomography (collimation 2×128×0.6mm, high pitch spiral data acquisition mode, 40-60ml contrast agents, radiation dose 3.5±0.9mSv). The following aortic annulus characteristics were determined by three independent observers: aortic annulus maximum, minimum and mean diameters (Dmax, Dmin, Dmean), eccentricity index (EI), effective aortic annulus diameter according to its circumference (Dcirc), effective aortic annulus diameter according to its area (Darea), distance from the aortic annulus plane to the left (LCA) and right coronary artery (RCA) ostia, maximum (DmaxAR) and minimum aortic root diameter (DminAR), maximum (DmaxSTJ) and minimum diameter of the sinotubular junction (DminSTJ). Subsequently, interobserver variabilities were assessed. RESULTS: Correlation between the three observers showed moderate to close agreement (between r=0.67 and r=0.97, all p<0.001). Mean differences (SE) between the three observers ranged from 0.07 (0.06)mm to 0.24 (0.07)mm for assessing the mean AA diameter (Dmean), from 0.28 (0.04)mm to 0.60 (0.06)mm for determining the effective AA diameter derived from the annulus area (Darea) and from 0.03 (0.07)mm to 0.07 (0.11)mm derived from the AA perimeter (Dcirc). For measurements of LCA and RCA distances to the AA level, mean interobserver differences (SE) ranged from 0.36 (0.07)mm to 0.76 (0.09)mm and from 0.15 (0.06)mm to 0.45 (0.11)mm. CONCLUSION: Computed tomography provides reproducible measurements of the aortic annulus and root geometry in patients scheduled for TAVI. The perimeter-derived aortic annulus diameter shows the lowest interobserver differences. Interobserver variabilities in prosthesis size recommendation were further reduced, if all three sizing methods were considered and stated as a "consensus result".