Despite being considered an important anatomical parameter directly related to neuronal density, cortical thickness is not routinely assessed in studies of the human brain in vivo. This paucity has been largely due to the size and convoluted shape of the human cortex, which has made it difficult to develop automated algorithms that can measure cortical thickness efficiently and reliably. Since the development of such an algorithm by Fischl and Dale in 2000, the number of studies investigating the relationship between cortical thickness and other physiological parameters in the brain has been on the rise. There have been no studies however that have validated cortical asymmetry against known vascular anatomy. To this aim, using high-resolution MRI, we measured cortical thickness and volume in the primary motor (M1) and primary visual (V1) cortex in patients with unilateral, high-grade carotid occlusive disease (n = 29, age = 74 ± 10 years). These regions were selected based on the hypothesis that there will be thinning of the cortical thickness of M1 in the territory supplied by the occluded carotid artery, whereas V1 will show no asymmetry since its blood supply is provided by unaffected posterior arteries. To test for an effect of handedness, cortical thickness and volume were also measured in healthy volunteers (n = 8, age = 37 ± 13 years). In patients, we found thinner cortex in M1 on the occluded side (mean = 2.07 ± 0.19 mm vs 2.15 ± 0.20 mm, p = 0.0008) but no hemispheric difference in V1 (1.80 ± 0.17 mm in occluded vs 1.78 ± 0.16 mm in unoccluded, p = 0.31). Although the mean cortical volume of M1 in the occluded hemisphere was also lower, the difference did not reach statistical significance (p = 0.09). Similarly, in healthy controls, the results showed no hemispheric asymmetry in either cortical thickness or volume in either region (p > 0.1). To test for an orientation bias in the method, the analysis was repeated with images flipped from neurological to radiological orientation. While the algorithm did not yield identical results for the two orientations, the effect did not alter the findings of the study. These results provide a method for within-subject validation of a pathophysiological effect of carotid occlusive disease on the human cortex and warrant further investigation for underlying mechanisms.
The purpose of this study is to validate the accuracy of abundance map reference data (AMRD) for three airborne imaging spectrometer (IS) scenes. AMRD refers to reference data maps (“ground truth”) that are specifically designed to quantitatively assess the performance of spectral unmixing algorithms. While classification algorithms typically label whole pixels as belonging to certain ground cover classes, spectral unmixing allows pixels to be composed of fractions or abundances of each class. The AMRD validated in this paper were generated using our previously-proposed remotely-sensed reference data (RSRD) technique, which spatially aggregates the results of standard classification or unmixing algorithms from fine spatial-scale IS data to produce AMRD for co-located coarse-scale IS data. Validation of the three scenes was accomplished by estimating AMRD in 51 randomly-selected 10 m×10 m plots, using seven independent methods and observers. These independent estimates included field surveys by two observers, imagery analysis by two observers and RSRD by three algorithms. Results indicated statistically-significant differences between all versions of AMRD. Even AMRD from our two field surveys were significantly different for two of the four ground cover classes. These results suggest that all forms of reference data require validation prior to use in assessing the performance of classification and/or unmixing algorithms. Given the significant differences between the independent versions of AMRD, we propose that the mean of all (MOA) versions of reference data for each plot and class is most likely to represent true abundances. Our independent versions of AMRD were compared to MOA to characterize error and uncertainty. Best case results were achieved by a version of imagery analysis, which had mean coverage area differences of 2.0%, with a standard deviation of 5.6%. One of the RSRD algorithms was nearly as accurate, achieving mean differences of 3.0%, with a standard deviation of 6.3%. Further analysis of statistical equivalence yielded an overall zone of equivalence between for this version of RSRD. The relative accuracy of RSRD methods is promising, given their potential to efficiently generate scene-wide AMRD. These results provide the first known validated abundance level reference data for airborne IS data.
15 June 2020 in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
Understanding and exploiting topographical data via standard machine learning techniques is challenging, mainly due to the large dynamic range of values present in elevation data and the lack of direct relationships between anthropogenic phenomena and topography, when considering topographic–geology couplings, for instance. Here we consider the first hurdle, dynamic range, in an effort to apply Convolutional Neural Network (CNN) approaches for prediction of human activity. CNN for learning 3-D elevation data relies on data normalization approaches, which only consider locally available points, thereby discarding contextual information and eliminating global contrast cues. We present a fully invertible and data-driven global partitioning elevation normalization (GPEN) preprocessing technique, which is intended to ameliorate the impact of limited data dynamic range. Global elevation populations are derived and used to formulate a distribution, which is used to adopt a partitioning scheme to remap all values according to global occurrence frequency, while preserving partition contrast. Using USGS 3-D Elevation Project and Microsoft building footprint data, we conduct a binary classification experiment predicting building footprint presence from elevation data, with and without a global remapping using the SegNet convolutional encoder-decoder model. The results of the experiment show more rapid model convergence, reduced regionalization errors, and enhanced classification metrics when compared to standard normalization preprocessing techniques. GPEN demonstrates performance over 10% higher than the next best conventional preprocessing method, with a mean overall accuracy of 94.76%. GPEN may show promise as an alternative normalization for deep learning with topological data.
1 May 2020 in IBM Journal of Research and Development
Architectural and hyperparameter design choices can influence deep-learner (DL) model fidelity but can also be affected by malformed training and validation data. However, practitioners may spend significant time refining layers and hyperparameters before discovering that distorted training data were impeding the training progress. We found that an evolutionary algorithm (EA) can be used to troubleshoot this kind of DL problem. An EA evaluated thousands of DL configurations on Summit that yielded no overall improvement in DL performance, which suggested problems with the training and validation data. We suspected that contrast limited adaptive histogram equalization enhancement that was applied to previously generated digital surface models, for which we were training DLs to find errors, had damaged the training data. Subsequent runs with an alternative global normalization yielded significantly improved DL performance. However, the DL intersection over unions still exhibited consistent subpar performance, which suggested further problems with the training data and DL approach. Nonetheless, we were able to diagnose this problem within a 12-hour span via Summit runs, which prevented several weeks of unproductive trial-and-error DL configuration refinement and allowed for a more timely convergence on an ultimately viable solution.
1 October 2020 in Photogrammetric Engineering & Remote Sensing
Terrestrial laser scanning systems are characterized by a sampling pattern which varies in point density across the hemisphere. Additionally, close objects are over-sampled relative to objects that are farther away. These two effects compound to potentially bias the three-dimensional statistics of measured scenes. Previous methods of sampling have resulted in a loss of structural coherence. In this article, a method of sampling is proposed to optimally sample points while preserving the structure of a scene. Points are sampled along a spherical coordinate system, with probabilities modulated by elevation angle and squared distance from the origin. The proposed approach is validated through visual comparison and stem-volume assessment in a challenging mangrove forest in Micronesia. Compared to several well-known sampling techniques, the proposed approach reduces sampling bias and shows strong performance in stem-reconstruction measurement. The proposed sampling method matched or exceeded the stem-volume measurement accuracy across a variety of tested decimation levels. On average it achieved 3.0% higher accuracy at estimating stem volume than the closest competitor. This approach shows promise for improving the evaluation of terrestrial laser-scanning data in complex scenes.