Rahul Dodhia

Digital infrastructure is crucial for economic opportunities, social inclusion, and civic participation, but benefits remain unevenly distributed. Existing measures, such as computer ownership and broadband adoption, do not capture the depth and variety of digital use. We develop measures using Microsoft telemetry collected during operating system updates from 40 million Windows devices in U.S. households. These measures provide greater coverage and granularity than existing commercial or government surveys. Device representations at the ZIP-code level reveal substantial variation in usage, including an urban-rural divide and disparities within narrowly defined Metropolitan Statistical Areas. Income and education correlate with these differences. These effects are large and suggest that digital usage gaps extend beyond the local availability of essential IT infrastructure. The findings call for interventions that complement infrastructure access with support for technology use and skills development. The resulting indices are publicly available to support future research on digital inequality.

Background/Objectives: The training of diagnostic pancreatic pathologists is largely limited by the diversity of available pathology images, especially those of rare diseases or conditions. Methods: Using a cohort of seven pancreatic neoplasms, we developed an iterative, pathologist-in-the-loop workflow integrating scalable tile pruning, generative-model optimization, and postprocessing truncation. Two pathologists evaluated synthetic-image quality and subtype representation on a 0–3 scale; the independent pathologist was blinded to image source and truncation condition and also rated curated real training tiles. Results: Untruncated images had lower class-balanced FID than per-class-truncated images (6.64 versus 29.08) and higher recall and coverage, whereas truncation increased precision. The independent pathologist rated per-class-truncated images higher than matched non-truncated images (mean difference 0.80, bootstrap 95% CI 0.57–1.03). Quadratic-weighted Cohen’s κ was 0.599 (95% CI 0.489–0.691); after grouping ratings as 0–1 versus 2–3, raw agreement was 80.7% (95% CI 75.0–86.4%). Conclusions: Per-class truncation improved independently rated image quality and subtype representation. Successful synthetic histology generation requires careful data curation, domain-specific oversight, and independent validation; clinical utility requires separate task-based evaluation.

Fence is a globally ubiquitous form of linear infrastructure that interrupts animal movement, compromises population fitness, alters ecological processes, and diminishes landscape connectivity. However, fences are also widely used as conservation and land-management tools to mitigate human–wildlife conflict. Despite their significant relevance to biodiversity conservation, they remain largely absent from regional to global biodiversity assessments, in contrast to other linear infrastructures such as roads and railways. This omission stems primarily from the absence of large-scale spatial datasets and scalable, consistent mapping methods. Existing efforts rely heavily on field surveys or manual interpretation of remote sensing imagery, both of which are time-consuming, labor-intensive, and difficult to scale, thereby limiting their broader applicability and long-term utility.Here, we introduce FenceMapper, an AI-driven framework for accurate and scalable fence detection using multi-scale segmentation of high-resolution remote sensing imagery. A local model extracts fine-scale fence features from small image patches, while a global refinement model improves structural continuity by incorporating broader contextual information and reducing fragmentation. Using 972 sampling sites containing 24,632 km of fences and more than 170,000 image patches across western U.S. rangelands, we trained and evaluated the framework and demonstrated robust performance: FenceMapper achieved 77% correctness, 75% completeness, and 76% quality under a 10 m tolerance, with highly consistent fence-length estimates (R² = 0.83). We then scaled the workflow to map 740,283 km of fences across the rangelands of the western United States, producing the most comprehensive sub-continental fence dataset to date.FenceMapper and the resulting open-access dataset provide a critical missing layer for biodiversity monitoring. For example, ecologists can pair spatial fence location with animal tracking or occurrence data to assess how fences alter animal behavior and influence community dynamics. Large-scale spatial fence density complements current landscape connectivity assessment, hence provides critical information for landscape-level conservation prioritization across diverse ecological settings. This work demonstrates how remote sensing and AI can address key gaps in global biodiversity observation systems and strengthen the integration of structural landscape information into conservation planning and decision support.

Purpose: To develop SCOPE (Small-lesion COntextual Pancreatic Evaluator), a deep learning model designed to improve CT detection of small pancreatic lesions—pancreatic ductal adenocarcinoma (PDAC), pancreatic neuroendocrine tumors (PanNETs), and cystic lesions—by integrating voxel-level features with global context. Materials and Methods: This retrospective study used three independent datasets. A development cohort of 4,065 contrast-enhanced CT scans was used to train a deep neural network that performs pancreas, ductal, and lesion segmentation with an integrated classification head. A metamodel combined segmentation-derived and global contextual signals for case-level prediction. Performance was assessed on (1) an internal holdout test set ( n = 605), (2) an external multi-institutional PDAC dataset from the PANORAMA challenge ( n = 2,238), and (3) an expert-curated small-lesion reader study ( n = 200). Areas under the receiver operating characteristic curve (AUCs) were compared using DeLong test; sensitivities and specificities using McNemar’s test. Results: On the internal test set, SCOPE improved lesion-versus-normal AUC compared with the best segmentation baseline (0.974 [95% CI: 0.964, 0.984] vs 0.956; P = .006) and increased small-lesion sensitivity at 95% specificity (0.727 [95% CI: 0.653, 0.801] vs 0.600; P = .012). Performance gains were observed across lesion classes, with significant improvements for PDAC and PanNET detection. On the external dataset, SCOPE improved PDAC-versus-non-PDAC AUC (0.978 vs 0.861, P < .001) and achieved higher sensitivity at 90% and 95% specificity without retraining. For the small-lesion reader study, SCOPE achieved lesion-versus-normal AUC of 0.922 and performed within the range of subspecialty abdominal radiologists; SCOPE provided the correct diagnosis in 14.5% (29/200) of cases in which two or more readers were incorrect. Conclusion: SCOPE improves early detection of small, potentially curable, pancreatic lesions on CT by combining local segmentation and global pancreatic context. Its consistent performance across internal, external, and reader datasets supports potential use as a concurrent reader for earlier and more accurate pancreatic lesion detection.

The increasing accessibility of remotely sensed data and their potential to support large-scale decision-making have driven the development of deep learning models for many Earth Observation tasks. Traditionally, such models rely on large datasets. However, the common assumption that larger training datasets lead to better performance tends to overlook issues related to data redundancy, noise, and the computational cost of processing massive datasets. Effective solutions must therefore consider not only the quantity but also the quality of data. Towards this, in this paper, we introduce six basic core-set selection approaches – that rely on imagery only, labels only, or a combination of both – and investigate whether they can identify high-quality subsets of data capable of maintaining – or even surpassing – the performance achieved when using full datasets for remote sensing semantic segmentation. We benchmark such approaches against two traditional baselines on three widely used land-cover classification datasets (DFC2022, Vaihingen, and Potsdam) using two different architectures (SegFormer and U-Net), thus establishing a general baseline for future works. Our experiments show that all proposed methods consistently outperform the baselines across multiple subset sizes, with some approaches even selecting core sets that surpass training on all available data. Notably, on DFC2022, a selected subset comprising only 25% of the training data yields slightly higher SegFormer performance than training with the entire dataset. This result shows the importance and potential of data-centric learning for the remote sensing domain. The code is available at https://github.com/keillernogueira/data-centric-rs-classification/.

Background: Predicting whether an organ offer will be accepted for transplantation remains challenging for several reasons, including large offer volumes, highly imbalanced observations (more declines than acceptances), and lack of information about the human decision‐making process. Offer acceptance models are used for risk‐adjusted program evaluations and policy development, but there is a lack of literature on baselines and best practices for predictive applications. We compared a suite of machine learning models, feature sets, and sampling procedures to identify performance impacts when training offer acceptance prediction models. Methods: We evaluated several kidney offer acceptance models from logistic regression to gradient boosted trees that were trained on donor and candidate characteristics. We then selected the best‐performing model and augmented training data with additional features (e.g., distance from the closest airport to the transplant hospital) or additional sampling procedures (e.g., undersampling). Results: Compared to the baseline logistic regression model (average precision: 0.0645), the XGBoost model offered the best performance improvement over the baseline (average precision: 0.0907). Including transportation‐related features in the model further improved model performance (average precision: 0.0940); however, we did not observe substantial model performance differences based on the sampling procedure used. Conclusions: Leveraging advanced machine learning models and incorporating nonclinical datapoints (like transportation distances) can improve transplant organ offer acceptance prediction models. However, we observed steep tradeoffs between precision and recall as captured in the low average precision scores despite deceptively high AUROCs (baseline AUROC 0.832). Our findings suggest that even the best‐performing models would not provide clear, equitable benefits over existing allocation policies. More research is needed before these models are practical for clinical implementation.

Background: Deep learning–based pancreas segmentation in CT has advanced rapidly yet remains evaluated primarily with mean overlap metrics that fail to capture robustness—defined as the proportion of cases reaching human-level performance. Models performing well on mean Dice or surface metrics can still fail unpredictably across scanners or anatomies. Because early detection and quantitative biomarkers rely on consistent segmentation, robustness is critical for clinical deployment. Purpose: To systematically evaluate the robustness of deep learning models for pancreas segmentation relative to human readers and to investigate an active learning strategy to improve reliability. Materials and Methods: We retrospectively assembled 903 venous-phase CT scans from patients with presumed normal-appearing pancreases and without known pancreatic disease (2005-2023), split into 803 for training/validation and 100 healthy test cases. Each test case had 4 independent human segmentations. Inter-reader variability on this healthy-only test set defined the empirical human distribution, providing an upper-bound estimate of robustness. We introduced a Fractional Threshold (FT) metric, measuring the proportion of model predictions exceeding the minimum human performance. Robustness was assessed across models trained from scratch, fine-tuned, or pretrained, including both normal and abnormal cases. An active learning approach identified high-uncertainty predictions for human revision. Statistical comparisons were performed using the Wilcoxon signed-rank and proportions Z-tests. Results: The best model, a 3-dimensional U-Net trained from scratch, achieved a Dice Similarity Coefficient (DSC) of 0.88 ± 0.04 and Normalized Surface Dice (NSD) of 0.77 ± 0.09, approaching human-level segmentation (DSC = 0.89 ± 0.03; NSD = 0.75 ± 0.07). However, FT for DSC and NSD remained lower than human performance in most cases, indicating persistent model variability. Human-in-the-loop revision of acquisition-flagged outliers increased FT to 0.99, with an average time of 1.54 minutes per case, corresponding to a 23-fold workload reduction. Conclusion: Automated pancreas segmentation reduces workload but remains constrained by tail-case failures. Active learning enhances model reliability, bridging the gap between artificial intelligence and human-level performance.