Search NASA⌕ Search

SEARCH · Search NASA

Results for “image coding”

Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 1,333 records · Page 74

INVESTIGATION OF BREAKDOWN INDUCED SURFACE DAMAGE ON 805 MHZ PILLBOX CAVITY INTERIOR SURFACES

The MuCool Test Area (MTA) at Fermilab is a facility to develop the technology required for ionization cooling for a future Muon Collider and/or Neutrino Factory. As part of this research program, we have tested two 805 MHz vacuum RF cavities in a multi-Tesla magnetic field to study the effects of the static magnetic field on the cavity operation. This study gives useful information on field emitters in the cavity, dark current, surface conditioning, breakdown mechanisms and material properties of the cavity. All these factors determine the maximum accelerating gradient in the cavity. This paper discusses the image processing technique for quantitative estimation of spark damage spot distribution on cavity interior surfaces. The distribution is compared with the electric field distribution predicted by a computer code calculation. The local spark density is proportional to probability of surface breakdown and shows a power law dependence on the maximum electric field (E). This E dependence is consistent with the dark current calculated from the Fowler-Nordheim equation.

43 PARTICLE ACCELERATORS↗

A Compact, Multiangle Electron Spectrometer for Ultraintense Laser-Plasma Interaction Experiments

Experiments on the multiterawatt (MTW) laser at the Laboratory for Laser Energetics will study the effect of the focal-spot shape on the forward acceleration and collimation of electrons. A compact electron spectrometer has been developed to record the energy spectra of electrons ejected in the interaction of the laser at multiple angular locations simultaneously. The modular system with replaceable magnets provides an adjustable energy band, currently 0.2-6 MeV. The detector is an array of imaging plates. The device is designed to operate in the high-noise environment (bremsstrahlung and Compton x rays, gamma rays, and scattered electrons), while being compact enough to fit in the 30 cm radius MTW target chamber. The detector geometry and shielding were optimized with the particle/radiation transport code GEANT4. Calibration was performed with beta sources. The required dynamic range, sensitivity, and resolution were confirmed with initial MTW experimental data.

Gotchev, O V↗

An image filtering technique for SPIDER visible tomography

The tomographic diagnostic developed for the beam generated in the SPIDER facility (100 keV, 50 A prototype negative ion source of ITER neutral beam injector) will characterize the two-dimensional particle density distribution of the beam. The simulations described in the paper show that instrumental noise has a large influence on the maximum achievable resolution of the diagnostic. To reduce its impact on beam pattern reconstruction, a filtering technique has been adapted and implemented in the tomography code. This technique is applied to the simulated tomographic reconstruction of the SPIDER beam, and the main results are reported.

70 PLASMA PHYSICS AND FUSION TECHNOLOGY↗

Autoradiography Using OSL for Monitoring Warhead Dismantlement

The use of radiation imaging techniques to provide increased confidence in a dismantlement verification regime is still under investigation. Currently, radiation measurements can be used to confirm attributes such as the presence, quality, and mass of containerized nuclear material. However, imaging techniques can be used to confirm the shape, size, and/or symmetry of the object being interrogated. Imaging can be used as a complementary technique to radiation measurements to confirm a declaration or as a chain of custody measure to confirm that an item has not been tampered with. As with other attribute measurements, the image data must be protected behind an information barrier and analyzed to give an unclassified result when compared to an agreed upon threshold. This paper will discuss the use of Optically Stimulated Luminescence to perform autoradiography of sources using coded apertures to differentiate between point sources and various distributed sources. The coded aperture chosen for this application is based on modified uniform redundant arrays (mURAs) which resulted in the design and manufacture of an efficient, high-resolution mURA mask/antimask system. The potential use of the autoradiography technique with the coded aperture as an intrinsic information barrier and the required protection of the mask and unfolding algorithm will also be discussed. This project has been supported by the US Department of Energy’s National Nuclear Security Administration’s Office of Dismantlement and Transparency (DOE/NNSA/NA-241).

11 NUCLEAR FUEL CYCLE AND FUEL MATERIALS↗

Monte Carlo and Analytical Calculation of Lateral Deflection of Proton Beams in Homogeneous Targets

Proton radiation therapy is a precise form of radiation therapy, but the avoidance of damage to critical normal tissues and the prevention of geographical tumor misses require accurate knowledge of the dose delivered to the patient and the verification of his position demand a precise imaging technique. In proton therapy facilities, the X-ray Computed Tomography (xCT) is the preferred technique for the planning treatment of patients. This situation has been changing nowadays with the development of proton accelerators for health care and the increase in the number of treated patients. In fact, protons could be more efficient than xCT for this task. One essential difficulty in pCT image reconstruction systems came from the scattering of the protons inside the target due to the numerous small-angle deflections by nuclear Coulomb fields. The purpose of this study is the comparison of an analytical formulation for the determination of beam lateral deflection, based on Moliere's theory and Rutherford scattering with Monte Carlo calculations by SRIM 2008 and MCNPX codes.

62 RADIOLOGY AND NUCLEAR MEDICINE↗

Photometric redshift analysis in the Dark Energy Survey Science Verification data

In this study, we present results from a study of the photometric redshift performance of the Dark Energy Survey (DES), using the early data from a Science Verification period of observations in late 2012 and early 2013 that provided science-quality images for almost 200 sq. deg. at the nominal depth of the survey. We assess the photometric redshift (photo-z) performance using about 15 000 galaxies with spectroscopic redshifts available from other surveys. These galaxies are used, in different configurations, as a calibration sample, and photo-z's are obtained and studied using most of the existing photo-z codes. A weighting method in a multidimensional colour–magnitude space is applied to the spectroscopic sample in order to evaluate the photo-z performance with sets that mimic the full DES photometric sample, which is on average significantly deeper than the calibration sample due to the limited depth of spectroscopic surveys. In addition, empirical photo-z methods using, for instance, artificial neural networks or random forests, yield the best performance in the tests, achieving core photo-z resolutions σ68 ~ 0.08. Moreover, the results from most of the codes, including template-fitting methods, comfortably meet the DES requirements on photo-z performance, therefore, providing an excellent precedent for future DES data sets.

46 INSTRUMENTATION RELATED TO NUCLEAR SCIENCE AND ↗

OpenSolar: Promoting the Openness and Accessibility of Diverse Public Solar Datasets

Observational solar data is the foundation of data-driven research in solar power grid integration and power system operations. Compared to other fields in data science, the openness and accessibility of solar data is lacking, which prevents solar data science from catching up with the emerging trends of data science (e.g., deep learning). In this paper, OpenSolar, a package with both R and Python versions, is developed to enhance the openness and accessibility of publicly available solar datasets. The OpenSolar package provides access to multiple types of solar data, primarily from four datasets: (1) the National Renewable Energy Laboratory (NREL) Solar Power Data for Integration Studies dataset, (2) the NREL Solar Radiation Research Laboratory dataset, (3) the Sheffield Solar-Microgen database, and (4) the Dataport database. Unlike other open solar datasets that only contain meteorological data, the four datasets in the OpenSolar package also contain behind-the-meter power data, sky images, and solar power data with satisfactory temporal and spatial resolution and coverage. The overview, quality-control methods, and potential usage of the datasets, in conjunction with sample code implementing the OpenSolar functions, are described in this paper. The package is expected to assist in bridging the gaps between the research fields of solar energy, power systems, and data science.

14 SOLAR ENERGY↗

An evolution of technologies and applications of gamma imagers in the nuclear cycle industry

The tracking of radiation contamination and distribution has become a high priority in the nuclear cycle industry in order to respect the ALARA principle which is a main challenge during decontamination and dismantling activities. To support this need, AREVA/CANBERRA and CEA LIST have been actively carrying out research and development on a gamma-radiation imager. In this paper we will present the new generation of gamma camera, called GAMPIX. This system is based on the Timepix chip, hybridized with a CdTe substrate. A coded mask could be used in order to increase the sensitivity of the camera. Moreover, due to the USB connection with a standard computer, this gamma camera is immediately operational and user-friendly. The final system is a very compact gamma camera (global weight is less than 1 kg without any shielding) which could be used as a hand-held device for radioprotection purposes. In this article, we present the main characteristics of this new generation of gamma camera and we expose experimental results obtained during in situ measurements. Even though we present preliminary results the final product is under industrialization phase to address various applications specifications. (authors)

12 MANAGEMENT OF RADIOACTIVE AND NON-RADIOACTIVE W↗

Automation Technologies Impact on the Work Process of Nuclear Power Plants

The current cost associated with the work process of nuclear power plants necessities modernizing the work process by augmenting it with automation technologies. This work was performed to assess various automation technologies and the means by which they assimilate into the work process in order to make gains in cost savings in modern nuclear power plants. The goal was to identify key informational items that would be useful to members of industry who are undertaking the large effort to reduce costs through automation by identifying technology maturity levels, recent research, and an analysis of cost impacts of the work process. The technologies cost saving was categorized into low, medium, and high cost savings. Through this effort, a substantive literature review and analysis of recent technological and market trends was performed. The key areas of cost impact were identified through communication with users in industry from operations and maintenance organizations. The results were assembled into a list of work process steps and technologies that are ready be implemented or researched for the automating aspects of that step. This effort aims to present to utilities opportunities of cost saving that are achievable in the short term, and gaps for research organizations and vendors to consider for further research and development. The results of this effort generated a list of technologies that fit each of the three cost saving categories. Electronic work packages remain a high cost saving technology, but data mining and integration, smart scheduling, and dynamic risk models are also indicated to yield a high cost savings. Video applications, spatial mapping, and image processing and identification technologies appear to provide solid cost savings in the medium category, and rounding out the low cost saving areas are deployable technologies such as bar codes, communication tools, and electronic tags.

46 INSTRUMENTATION RELATED TO NUCLEAR SCIENCE AND ↗

Deep-learning-based image registration for nano-resolution tomographic reconstruction

Nano-resolution full-field transmission X-ray microscopy has been successfully applied to a wide range of research fields thanks to its capability of non-destructively reconstructing the 3D structure with high resolution. Due to constraints in the practical implementations, the nano-tomography data is often associated with a random image jitter, resulting from imperfections in the hardware setup. Without a proper image registration process prior to the reconstruction, the quality of the result will be compromised. Here a deep-learning-based image jitter correction method is presented, which registers the projective images with high efficiency and accuracy, facilitating a high-quality tomographic reconstruction. This development is demonstrated and validated using synthetic and experimental datasets. We report the method is effective and readily applicable to a broad range of applications. Together with this paper, the source code is published and adoptions and improvements from our colleagues in this field are welcomed.

deep learning↗

Hohlraum Designs for High Velocity Implosions on NIF

In this paper, we compare experimental shock and capsule trajectories to design calculations using the radiation-hydrodynamics code HYDRA. The measured trajectories from surrogate ignition targets are consistent with reducing the x-ray flux on the capsule by about 85%. A new method of extracting the radiation temperature as seen by the capsule from x-ray intensity and image data shows that about half of the apparent 15% flux deficit in the data with respect to the simulations can be explained by HYDRA overestimating the x-ray flux on the capsule. The National Ignition Campaign (NIC) point-design target is designed to reach a peak fuel-layer velocity of 370 km/s by ablating 90% of its plastic (CH) ablator. The 192-beam National Ignition Facility laser drives a gold hohlraum to a radiation temperature (T{sub RAD}) of 300 eV with a 20 ns-long, 420 TW, 1.3 MJ laser pulse. The hohlraum x-rays couple to the CH ablator in order to apply the required pressure to the outside of the capsule. In this paper, we compare experimental measurements of the hohlraum T{sub RAD} and the implosion trajectory with design calculations using the code hydra. The measured radial positions of the leading shock wave and the unablated shell are consistent with simulations in which the x-ray flux on the capsule is artificially reduced by 85%. We describe a new method of inferring the T{sub RAD} seen by the capsule from time-dependent x-ray intensity data and static x-ray images. This analysis shows that hydra overestimates the x-ray flux incident on the capsule by {approx}8%.

70 PLASMA PHYSICS AND FUSION TECHNOLOGY↗

Investigation of ion and electron heat transport of high- T e ECH heated discharges in the large helical device

An analysis of the radial electric field and heat transport, both for ions and electrons, is presented for a high-${{T}_{\text{e}}}$ electron cyclotron heated (ECH) discharge on the large helical device (LHD). Transport analysis is done using the task3d transport suite utilizing experimentally measured profiles for both ions and electrons. Ion temperature and perpendicular flow profiles are measured using the recently installed x-ray imaging crystal spectrometer diagnostic (XICS), while electron temperature and density profiles are measured using Thomson scattering. The analysis also includes calculated ECH power deposition profiles as determined through the travis ray-tracing code. This is the first time on LHD that this type of integrated transport analysis with measured ion temperature profiles has been performed without NBI, allowing the heat transport properties of plasmas with only ECH heating to be more clearly examined. For this study, a plasma discharge is chosen which develops a high central electron temperature (${{T}_{\text{eo}}}=9$ keV) at moderately low densities (${{n}_{\text{eo}}}=1.5\times {{10}^{19}}$ m-3). The experimentally determined transport properties from task3d are compared to neoclassical predictions as calculated by the gsrake and fortec-3d codes. The predicted electron fluxes are seen to be an order of magnitude less than the measured fluxes, indicating that electron transport is largely anomalous, while the neoclassical and measured ion heat fluxes are of the same magnitude. Neoclassical predictions of a strong positive ambipolar electric field (${{E}_{\text{r}}}$ ) in the plasma core are validated through comparisons to perpendicular flow measurements from the XICS diagnostic. Furthermore, this provides confidence that the predictions are producing physically meaningful results for the particle fluxes and radial electric field, which are a key component in correctly predicting plasma confinement.

70 PLASMA PHYSICS AND FUSION TECHNOLOGY↗

MO-DE-210-05: Improved Accuracy of Liver Feature Motion Estimation in B-Mode Ultrasound for Image-Guided Radiation Therapy

Purpose: In similarity-measure based motion estimation incremental tracking (or template update) is challenging due to quantization, bias and accumulation of tracking errors. A method is presented which aims to improve the accuracy of incrementally tracked liver feature motion in long ultrasound sequences. Methods: Liver ultrasound data from five healthy volunteers under free breathing were used (15 to 17 Hz imaging rate, 2.9 to 5.5 minutes in length). A normalised cross-correlation template matching algorithm was implemented to estimate tissue motion. Blood vessel motion was manually annotated for comparison with three tracking code implementations: (i) naive incremental tracking (IT), (ii) IT plus a similarity threshold (ST) template-update method and (iii) ST coupled with a prediction-based state observer, known as the alpha-beta filter (ABST). Results: The ABST method produced substantial improvements in vessel tracking accuracy for two-dimensional vessel motion ranging from 7.9 mm to 40.4 mm (with mean respiratory period: 4.0 ± 1.1 s). The mean and 95% tracking errors were 1.6 mm and 1.4 mm, respectively (compared to 6.2 mm and 9.1 mm, respectively for naive incremental tracking). Conclusions: High confidence in the output motion estimation data is required for ultrasound-based motion estimation for radiation therapy beam tracking and gating. The method presented has potential for monitoring liver vessel translational motion in high frame rate B-mode data with the required accuracy. This work is support by Cancer Research UK Programme Grant C33589/A19727.

60 APPLIED LIFE SCIENCES↗

Kilovoltage Imaging Doses in the Radiotherapy of Pediatric Cancer Patients

Purpose: To investigate doses induced by kilovoltage cone-beam computed tomography (kVCBCT) to pediatric cancer patients undergoing radiotherapy, as well as strategies for dose reduction. Methods and Materials: An EGS4 Monte Carlo code was used to calculate three-dimensional dose deposition due to kVCBCT on 4 pediatric cancer patients. Absorbed doses to various organs were analyzed for both half-fan and full-fan modes. Clinical conditions, such as distance from organ at risk (OAR) to CBCT field border, kV peak energy, and testicular shielding, were studied. Results: The mean doses induced by one CBCT scan operated at 125 kV in half-fan mode to testes, liver, kidneys, femoral heads, spinal cord, brain, eyes, lens, and optical nerves were 2.9, 4.7, 7.7, 10.5, 8.8, 7.6, 7.7, 7.8, and 7.2 cGy, respectively. Increasing the distances from OARs to CBCT field border greatly reduced the doses to OARs, ranging from 33% reduction for spinal cord to 2300% reduction for testes. As photon beam energy increased from 60 to 125 kV, the dose increase due to kVCBCT ranged from 170% for lens to 460% for brain and spinal cord. A testicular shielding made of 1-cm cerrobend could reduce CBCT doses down to 31%, 51%, 68%, and 82%, respectively, for 60, 80, 100, and 125 kV when the testes lay within the CBCT field. Conclusions: Generally speaking, kVCBCT deposits much larger doses to critical structures in children than in adults, usually by a factor of 2 to 3. Increasing the distances from OARs to CBCT field border greatly reduces doses to OARs. Depending on OARs, kVCBCT-induced doses increase linearly or exponentially with photon beam energy. Testicular shielding works more efficiently at lower kV energies. On the basis of our study, it is essential to choose an appropriate scanning protocol when kVCBCT is applied to pediatric cancer patients routinely.

62 RADIOLOGY AND NUCLEAR MEDICINE↗

SU-F-J-08: Quantitative SPECT Imaging of Ra-223 in a Phantom

Purpose: Ra-223 therapy of prostate cancer bone metastases is being used to treat patients routinely. However, the absorbed dose distribution at the macroscopic and microscopic scales remains elusive, due to the inability to image the small activities injected. Accurate activity quantification through imaging is essential to calculate the absorbed dose in organs and sub-units in radiopharmaceutical therapy, enabling personalized absorbed dose-based treatment planning methodologies and more effective and optimal treatments. Methods: A 22 cm diameter by 20 cm long cylindrical phantom, containing a 3.52 cm diameter sphere, was used. A total of 2.01 MBq of Ra-223 was placed in the phantom with 177.6 kBq in the sphere. Images were acquired on a dual-head Siemens Symbia T16 gamma camera using three 20% full-width energy windows and centered at 84, 154, and 269 keV (120 projections, 360° rotation, 45 s per view). We have implemented reconstruction of Ra-223 SPECT projections using OS-EM (up to 20 iterations of 10 subsets) with compensation for attenuation using CT-based attenuation maps, collimator-detector response (CDR) (including septal penetration, scatter and Pb x-ray modeling), and scatter in the patient using the effective source scatter estimation (ESSE) method. The CDR functions and scatter kernels required for ESSE were computed using the SIMIND MC simulation code. All Ra-223 photon emissions as well as gamma rays from the daughters Rn-219 and Bi-211 were modeled. Results: The sensitivity of the camera in the three combined windows was 107.3 cps/MBq. The visual quality of the SPECT images was reasonably good and the activity in the sphere was 27% smaller than the true activity. This underestimation is likely due to partial volume effect. Conclusion: Absolute quantitative Ra-223 SPECT imaging is achievable with careful attention to compensate for image degrading factors and system calibration.

60 APPLIED LIFE SCIENCES↗

The Jefferson Lab High Power Broadband THz Facility

Jefferson Lab operates a broadband terahertz (THz) facility, whose light is generated by 300 fs FWHM 135 pC bunches of electrons at an energy of 125 MeV as they traverse a magnetic compression chicane prior to entering the optical cavity of the Jefferson Lab Free Electron Laser (FEL)[1-3]. The repetition rate of the electron bunches, and hence that of the light pulses, can be as high as 75 MHz continuous. Approximately 1 microJoule of energy is emitted for each pulse, which implies an average power of up to 75W, a peak power of several MW and electric fields of MV/m. Niche applications of this unique source are pump-probe dynamics, both in and out of equilibrium, and large-scale imaging. The beamline itself is designed to deliver an f/8 beam 2mm vertical × 3mm horizontal in size and 60% vertically polarized from a subtended angle of ~200 milliradians horizontal by ~135 milliradians vertical in the accelerator. The first mirror is copper-braid cooled. The philosophy of the optical design is to use a relay optics configuration to transport the beam via reflection off metal mirrors through a series of focal points, with the final one being at a diamond window. The optical scheme was developed using the Synchrotron Radiation Workshop (SRW)[4]. This code performs a full calculation of the electric field from a relativistic electron. It does not handle multiparticle coherent enhancement[5], but it is significantly different from all the other synchrotron radiation calculations because it retains a term called the Coulomb term in which the electric field falls off as 1/R2. In addition to calculating the electric field and the intensity, the code also allows one to propagate the field through the optical focusing elements and apertures. The beam exiting the final diamond window in the THz user lab passes into a series of vacuum chambers with considerable flexibility for attenuation, collimation, and/or beam splitting for imaging, spectroscopy and 2 photon pump-probe measurements. In this talk we will discuss the details of the philosophy, optical design and operating experience of this facility and also present ideas for discussion. We deeply appreciate the excellent support of our colleagues without which these facilities would not have been built and operated. This work was supported by the U.S. Dept. of Energy under contract DE-AC05-06-OR23177, and the Commonwealth of Virginia. The FEL facility is supported by the Office of Naval Research. 1. S. V. Benson et al, ?The VUV/IR/THz Free Electron Laser Program at Jefferson Lab?, Nucl. Instr. and Meth. A649 9 (2011). 2. G.L. Carr, M.C. Martin, W.R. McKinney, K. Jordan, G.R. Neil and G.P. Williams ?High Power Terahertz Radiation from Relativistic Electrons?, Nature 420 153-156 (2002). 3. J.M. Klopf et al. ?The Jefferson Lab high power THz user facility?, Nucl. Instrum. Methods A582, 114-116 (2007). 4. O. Chubar, P. Elleaume, "Accurate And Efficient Computation Of Synchrotron Radiation In The Near Field Region", proc. of the EPAC98 Conference, 22-26 June 1998, p.1177-1179. 5. C. J. Hirschmugl, M. Sagurton and G. P. Williams, ?Multiparticle coherence calculations for synchrotron radiation emission? Physical Review A44, 1316 (1991).

Klopf, John↗

MABAL: a Novel Deep-Learning Architecture for Machine-Assisted Bone Age Labeling

Bone age assessment (BAA) is a commonly performed diagnostic study in pediatric radiology to assess skeletal maturity. The most commonly utilized method for assessment of BAA is the Greulich and Pyle method (Pediatr Radiol 46.9:1269–1274, 2016; Arch Dis Child 81.2:172–173, 1999) atlas. The evaluation of BAA can be a tedious and time-consuming process for the radiologist. As such, several computer-assisted detection/diagnosis (CAD) methods have been proposed for automation of BAA. Classical CAD tools have traditionally relied on hard-coded algorithmic features for BAA which suffer from a variety of drawbacks. Recently, the advent and proliferation of convolutional neural networks (CNNs) has shown promise in a variety of medical imaging applications. There have been at least two published applications of using deep learning for evaluation of bone age (Med Image Anal 36:41–51, 2017; JDI 1–5, 2017). However, current implementations are limited by a combination of both architecture design and relatively small datasets. The purpose of this study is to demonstrate the benefits of a customized neural network algorithm carefully calibrated to the evaluation of bone age utilizing a relatively large institutional dataset. In doing so, this study will aim to show that advanced architectures can be successfully trained from scratch in the medical imaging domain and can generate results that outperform any existing proposed algorithm. The training data consisted of 10,289 images of different skeletal age examinations, 8909 from the hospital Picture Archiving and Communication System at our institution and 1383 from the public Digital Hand Atlas Database. The data was separated into four cohorts, one each for male and female children above the age of 8, and one each for male and female children below the age of 10. The testing set consisted of 20 radiographs of each 1-year-age cohort from 0 to 1 years to 14–15+ years, half male and half female. The testing set included left-hand radiographs done for bone age assessment, trauma evaluation without significant findings, and skeletal surveys. A 14 hidden layer-customized neural network was designed for this study. The network included several state of the art techniques including residual-style connections, inception layers, and spatial transformer layers. Data augmentation was applied to the network inputs to prevent overfitting. A linear regression output was utilized. Mean square error was used as the network loss function and mean absolute error (MAE) was utilized as the primary performance metric. MAE accuracies on the validation and test sets for young females were 0.654 and 0.561 respectively. For older females, validation and test accuracies were 0.662 and 0.497 respectively. For young males, validation and test accuracies were 0.649 and 0.585 respectively. Finally, for older males, validation and test set accuracies were 0.581 and 0.501 respectively. The female cohorts were trained for 900 epochs each and the male cohorts were trained for 600 epochs. An eightfold cross-validation set was employed for hyperparameter tuning. Test error was obtained after training on a full data set with the selected hyperparameters. Using our proposed customized neural network architecture on our large available data, we achieved an aggregate validation and test set mean absolute errors of 0.637 and 0.536 respectively. To date, this is the best published performance on utilizing deep learning for bone age assessment. Our results support our initial hypothesis that customized, purpose-built neural networks provide improved performance over networks derived from pre-trained imaging data sets. We build on that initial work by showing that the addition of state-of-the-art techniques such as residual connections and inception architecture further improves prediction accuracy. This is important because the current assumption for use of residual and/or inception architectures is that a large pre-trained network is required for successful implementation given the relatively small datasets in medical imaging. Instead we show that a small, customized architecture incorporating advanced CNN strategies can indeed be trained from scratch, yielding significant improvements in algorithm accuracy. It should be noted that for all four cohorts, testing error outperformed validation error. One reason for this is that our ground truth for our test set was obtained by averaging two pediatric radiologist reads compared to our training data for which only a single read was used. This suggests that despite relatively noisy training data, the algorithm could successfully model the variation between observers and generate estimates that are close to the expected ground truth.

62 RADIOLOGY AND NUCLEAR MEDICINE↗