Sensor simulation is a critical tool to address the gaps in real-world data for autonomous vehicle (AV) development. However, it is only effective if sensor models accurately reflect the physical world.
In previous posts, we detailed the validation process for camera and lidar models in NVIDIA DRIVE Sim. See Validating NVIDIA DRIVE Sim Camera Models and Validating NVIDIA DRIVE Sim Lidar Models. This post will cover radar, an essential sensor for obstacle detection and avoidance.
There are multiple ways to approach radar validation. You can compare how an AV stack trained on real-world data behaves when encountering synthetic radar data, for example. Or, you can compare synthetic radar data to its physical counterpart in real-world experiments.
Validating the model with an AV stack only evaluates its ability to the extent of triggering the AV function, which tests for a lower fidelity ceiling. For this reason, we will focus on the second approach.
We used radar sensors in the NVIDIA DRIVE Hyperion AV reference architecture, so developers building on NVIDIA DRIVE can easily transition between simulation and the real world. The sensors were mounted on a development vehicle (Figure 3). For this case, the front center radar (FCR) was the focus for evaluation.
Operating at a frequency of 77GHz, the radar under test included two scans: a near scan, with a wide FOV but limited range, and a far scan with extended range but a narrow FOV. Additionally, a 360 rotating lidar sensor (LD1) was mounted on top of the car to provide pseudo ground truth data.
Figure 6 depicts a top-down view of both real and simulated radar detections across all 1,211 high and low RCS corner reflector positions. We used this as a coherence check to start. Although we observed differences in FOV coverages above 80m, the overall coverage presented a noticeable similarity that is sufficient for the cross check.
The histograms in Figure 7 present the error distribution in range, azimuth, and RCS relative to the ground truth for both the high and low RCS corner reflectors, combined. Where applicable, we quantified the results by fitting a Gaussian distribution to the data. Results for the real radar are displayed on the left, while the DRIVE Sim data is shown on the right.
In scenarios where road objects are near each other (stationary vehicles under a bridge, pedestrians or motorcyclists next to a vehicle or guard rail, or two closely parked cars, for example) radars can encounter difficulties in distinguishing individual objects. For this reason, it is crucial to accurately simulate this characteristic, known as separation capability.
Results for all positions at 0 and 50m, and -45 and 20m, demonstrated a high degree of similarity between real and simulated. We observed a minor discrepancy at 0 and 50m where CR2 (50.5, 0). In this scenario, the real radar returned two detections instead of one in 10% of the scans.
As shown in Table 2, the results from the simulated and real-world sensors are largely in correlation. Significant deviations are noted at 0 and 50m where CR2 (50.5, 0). Furthermore, for 0 and 100m where CR1=CR2, the simulated radar returns two detections in 40% of scans, while real world never returns two detections.
Our simulation accurately replicated this behavior, as demonstrated by the alignment of the peaks in both the real and simulated data. Particularly, at a speed of 80kph, both the real and simulated radar exhibited similar velocity wrapping.
By validating accurate radar sensor behavior in simulated scenarios, we can improve system development efficiency, reduce dependence on costly and time-consuming real-world data collection, and enhance the safety and performance of AV systems.
Rodent species are widely used in the testing and approval of new radiopharmaceuticals, necessitating murine phantom models. As more therapy applications are being tested in animal models, calculating accurate dose estimates for the animals themselves becomes important to explain and control potential radiation toxicity or treatment efficacy. Historically, stylized and mathematically based models have been used for establishing doses to small animals. Recently, a series of anatomically realistic human phantoms was developed using body models based on nonuniform rational B-spline. Realistic digital mouse whole-body (MOBY) and rat whole-body (ROBY) phantoms were developed on the basis of the same NURBS technology and were used in this study to facilitate dose calculations in various species of rodents.
Methods: Voxel-based versions of scaled MOBY and ROBY models were used with the Vanderbilt multinode computing network (Advanced Computing Center for Research and Education), using geometry and tracking radiation transport codes to calculate specific absorbed fractions (SAFs) with internal photon and electron sources. Photon and electron SAFs were then calculated for relevant organs in all models.
Conclusion: The organ masses in the MOBY and ROBY models are in reasonable agreement with models presented by other investigators noting that considerable variation can occur between reported masses. Results consistent with those found by other investigators show that absorbed fractions for electrons for organ self-irradiation were significantly less than 1.0 at energies above 0.5 MeV, as expected for many of these small-sized organs, and measurable cross-irradiation was observed for many organ pairs for high-energy electrons (as would be emitted by nuclides such as (32)P, (90)Y, or (188)Re).
To combat these challenges and allow for accurate identification of differentially methylated loci, we present a novel approach to perform RNA methylAtion Differential Analysis in R (RADAR) for MeRIP-seq data. RADAR accounts for variation in pre-IP RNA and in post-IP read counts using different strategies. Specifically, RADAR uses gene-level read counts instead of peak-level read counts in the INPUT library as a robust measurement of the initial pre-IP RNA expression level. In addition, RADAR uses a flexible Poisson random effect model to accommodate over-dispersion in the post-IP read counts due to variability of biological replicates and noise introduced in the immunoprecipitation process. This generalized linear model framework enables incorporation of covariates in complex study designs.
We benchmarked the performance of RADAR with alternative methods on simulated data by different data generating models. We showed RADAR achieved higher sensitivity and specificity compared to existing alternative methods. We also demonstrated the performance of RADAR on real MeRIP-seq data by applying it to four high-quality m6A-meRIP-seq (aka m6A-seq) datasets generated by us and others, including an ovarian cancer dataset (GSE 119168) consisting of 7 normal fallopian tube tissue from healthy individuals and 6 metastatic omental tumors, a type 2 diabetes (T2D, GSE 120024) dataset consisting of human islets from 8 type II diabetes patients and 7 healthy control patients with samples being processed in three batches due to different sample acquisition times, a mouse liver (GSE 119490) dataset consisting of mouse liver from 4 wild type mice and 4 METTL14 knockout mice, and a mouse brain (GSE 113781) dataset consisting of 7 mouse cortex samples of stress-exposed mice and 7 from control mice. We showed that our approach can accommodate distinct study designs and led to more sensitive and reproducible DM locus identification than possible alternatives.
We then repeated simulation studies using the QNB model. Instead of setting the variances of INPUT and IP libraries equal as presented in the QNB paper, we let the variance of IP read count be larger than that of INPUT. This setting better reflects our observation in the real data as extra noise can be introduced during immunoprecipitation process for IP reads generation (Additional file 1: Figure S4). In the simple case without covariates, RADAR exhibited the lowest empirical FDR (18.9% and 18.5%) despite slightly lower sensitivity comparing to other methods (73.5% and 82.3%) when the effect sizes were relatively large (for effect sizes of 0.75 and 1). QNB performed better when the effect size was small with 58.6% sensitivity and 15.6% FDR for an effect size of 0.5 (Fig. 2c). The results were consistent when we evaluated their performance with different FDR cutoffs. Overall, QNB performed slightly better than RADAR with an effect size of 0.5. RADAR achieved similar sensitivity but better calibrated FDR when effect sizes equal to 0.75 and 1 (Additional file 1: Figure S5C). In the model with covariates, RADAR exhibited the lowest empirical FDR, with 25.8%, 23.0%, and 22.5% at effect sizes of 0.5, 0.75, and 1, respectively, while other methods either failed to detect the signal or had a higher empirical FDR. Specifically, MeTDiff had sensitivity below 0.5% at varying effect sizes and QNB reached FDRs of 64.1%, 55.8%, and 50.5% for effect sizes of 0.5, 0.75, and 1, respectively, at an FDR cutoff of 10% (Fig. 2d). The advantage of RADAR over alternative methods hold in the difficult case at varying cutoffs (Additional file 1: Figure S5D). In summary, RADAR outperformed other existing methods in most scenarios, particularly when covariates were present.
In the T2D dataset, DMGs identified by RADAR were enriched in related pathways including insulin signaling pathways, type II diabetes mellitus, mTOR pathways, and AKT pathways (Additional file 1: Table S1), indicating a role that m6A might play in T2D. We further analyzed these DMGs in related pathways and found the methylome of insulin/IGF1-AKT-PDX1 signaling pathway been mostly hypo-methylated in T2D islets (Additional file 1: Figure S12). Impairment of this pathway resulting in downregulation of PDX1 has been recognized as a mechanism associated with T2D where PDX1 is a critical gene regulating β cell identity and cell cycle and promoting insulin secretion [21,22,23,24]. Indeed, follow-up experiment on a cell line model validated the role of m6A in tuning cell cycle and insulin secretion in β cells and animal model lacking methyltransferase Mettl14 in β cells recapitulated key T2D phenotypes (results presented in a separate manuscript, [25]). To summarize, RADAR-identified DMGs enabled us to pursue an in-depth analysis of the role that m6A methylation plays in T2D. On the contrary, due to the incapability to take sample acquisition batches as covariates, the alternative methods were underpowered to detect DM sites in T2D dataset and could not lead to any in-depth discovery of m6A biology in T2D islets. These examples suggest that MeRIP-seq followed by RADAR analysis could further advance functional studies of RNA modifications.
caa09b180b