Lawrence Hall

Objectives: Among surgically resected non-small cell lung cancer (NSCLC) patients with similar stage and histopathology, there is variability in patient outcomes highlighting the urgency of identifying biomarkers to predict recurrence. This study systematically develop a preoperative CT-based habitat-based radiomics classifier to predict recurrence in NSCLC. Methods: NSCLC patients with surgically resected stage IA–IIIA disease were randomly divided into a training and test cohorts. From pre-surgical CT images, tumor habitats were generated using two-level unsupervised clustering and then radiomic features were calculated from the intratumoral region and habitat subregions. Using ridge-regularized logistic regression, separate classifiers were developed to predict 3-year recurrence using intratumoral radiomics, habitat-based radiomics, and a combined model generated by a stacked learning framework. For each classifier, probability of recurrence was calculated for each patient then numerous statistical and machine learning approaches were utilized to stratify patients for recurrence-free survival.Results: The combined radiomics classifier yielded a superior AUC (0.82) compared to the intratumoral (AUC=0.75) and habitat radiomics (AUC=0.81) classifiers. When the classifiers were used to stratify high- versus low-risk patients utilizing a cut-point identified by decision tree analysis, high-risk patients were yielded the largest risk-of-recurrence (HR = 8.43; 95% CI 2.47–28.81) compared to the habitat (HR = 5.41; 95% CI 2.08–14.09) and intratumoral radiomics (HR = 3.54; 95% CI 1.45–8.66) models. SHAP analyses indicated that habitats contributed most strongly to recurrence prediction.Conclusions: This study revealed that habitat-based radiomics provided superior statistical performance than intratumoral radiomics for predicting recurrence in NSCLC.

The Deepwater Horizon (DWH) oil spill in the NE Gulf of Mexico in 2010 was the largest accidental spill globally. During and after the spill, oil was incorporated into marine snow (marine‐oil‐snow or MOS) and sedimented to the seafloor, accounting for ∼21% of the oil released. We assessed the abundance, distribution, size, composition, shape, and fractal dimension of marine snow in relation to the oil spill and natural environmental conditions between May 2010 and August 2014 using SIPPER camera images. The Mississippi River outflow, which governs productivity in this region, was elevated during the oil spill resulting in a large phytoplankton bloom. Integrated marine snow abundance was significantly higher during the spill and particle size spectra showed a higher abundance of intermediate‐ and large‐sized particles during the oil spill compared to non‐spill years. Particle collision and sinking rates may be sensitive to particle composition, shape, and fractal dimension (surface area complexity). Identifiable aggregate components included chain diatoms and acantharian radiolarians and their spines. Almost all aggregates were elongated in all years. The geometric mean fractal dimension for particles in non‐oil spill years was 1.45 (range: 1.00–1.94), whereas particle fractal dimensions were significantly higher (mean: 1.49, 1.01–1.87, p = 0.03) during the oil spill, indicating that oil droplets in MOS may have decreased particle porosity compared to other years. Fractal dimensions also increased with particle size during all years. These data provide a baseline for model development, responding to future oil spills, and for investigations of biogeochemical cycles in this region.

There have been many promising developments in deep learning to identify degrees of malignancies in prostate cancer pathologies. Deep network models have been shown to be useful in identifying patterns in histology images assessed at different scales. Prostate pathological grade identification has been a challenge among clinical experts due to complex patterns on the whole slide level, for hematoxylin and eosin (H&E) stained samples. In this study, we identify primary patterns (Gleason) in small sections of the whole slide composed of uniform glandular patterns. We then follow sample selection methods that eliminate ambiguous regions or tiled-samples by confidence filtering. A pseudo-confidence is derived from the predicted output of the network, which is used as a quality indicator to consider the sample for discriminatory analysis. We provide further evidence that using highly calibrated confidence sample selection, these gland-level features on the prostate biopsy sections can discriminate degrees of malignancy following primary Gleason patterns. We used an optimized deep network (convolutional neural network, CNN) discriminating glandular regions with aggressive grades (Gleason 3 from 4) showed an accuracy of 0.68(0.04), F ₁ of 0.66(0.06) and AUC of 0.74(0.04). We further improve this result using confidence filtering, with a sample fraction of 0.35 (with a calibrated confidence of greater than 0.85), achieving an accuracy of 0.74 (0.08), F ₁ of 0.72 (0.12), and AUC 0.79 (0.08) averaged from holdout sets over multiple reshuffled experiments.

Machine learning models trained on computed tomography (CT) images are highly sensitive to variations in imaging acquisition parameters. Even subtle inconsistencies, often unnoticeable to human radiologists, can significantly degrade model accuracy. In clinical practice, datasets frequently exhibit heterogeneity due to variations in imaging protocols and scanner characteristics, which makes associated metadata a valuable but often underutilized resource for identifying sources of bias. To address this, we propose a novel unsupervised method that systematically identifies confounding and potentially confounding factors embedded in metadata. The key strengths of our method include automated detection of influential metadata attributes, minimal reliance on manual input, and the capability to proactively flag variables that could induce model drift post-deployment. Empirical evaluation in two distinct CT datasets demonstrates that controlling for factors identified by our method drastically improves model performance, increasing classification accuracy by 5 to 15% compared to datasets where these factors remain uncontrolled. These comparative results underscore the potential of our approach to substantially improve the robustness, consistency, and clinical applicability of radiomic machine learning models.

Activation functions govern the expressivity and stability of neural networks, yet existing comparisons remain largely heuristic. We propose a rigorous framework for their classification via a nine-dimensional integral signature S_sigma(phi), combining Gaussian propagation statistics (m1, g1, g2, m2, eta), asymptotic slopes (alpha_plus, alpha_minus), and regularity measures (TV(phi'), C(phi)). This taxonomy establishes well-posedness, affine reparameterization laws with bias, and closure under bounded slope variation. Dynamical analysis yields Lyapunov theorems with explicit descent constants and identifies variance stability regions through (m2', g2). From a kernel perspective, we derive dimension-free Hessian bounds and connect smoothness to bounded variation of phi'. Applying the framework, we classify eight standard activations (ReLU, leaky-ReLU, tanh, sigmoid, Swish, GELU, Mish, TeLU), proving sharp distinctions between saturating, linear-growth, and smooth families. Numerical Gauss-Hermite and Monte Carlo validation confirms theoretical predictions. Our framework provides principled design guidance, moving activation choice from trial-and-error to provable stability and kernel conditioning.

Synthetic sequence-controlled polymers promise to transform polymer science by combining the chemical versatility of synthetic polymers with the precise sequence-mediated functionality of biological proteins. However, design of these materials has proven extraordinarily challenging, because they lack the massive datasets of closely related evolved molecules that accelerate design of proteins. Here we report on a new Artifical Intelligence strategy to dramatically reduce the amount of data necessary to accelerate these materials' design. We focus on data connecting the repeat-unit-sequence of a \emph{compatibilizer} molecule to its ability to reduce the interfacial tension between distinct polymer domains. The optimal sequence of these molecules, which are essential for applications such as mixed-waste polymer recycling, depends strongly on variables such as concentration and chemical details of the polymer. With current methods, this would demand an entirely distinct dataset to enable design at each condition. Here we show that a deep neural network trained on low-fidelity data for sequence/interfacial tension relations at one set of conditions can be rapidly tuned to make higher-fidelity predictions at a distinct set of conditions, requiring far less data that would ordinarily be needed. This priming-and-tuning approach should allow a single low-fidelity parent dataset to dramatically accelerate prediction and design in an entire constellation of related systems. In the long run, it may also provide an approach to bootstrapping quantitative atomistic design with AI insights from fast, coarse simulations.

Large language models (LLMs) have shown progress and promise in diverse applications ranging from the medical field to chat bots. Developing LLMs requires a large corpus of data and significant computation resources to achieve efficient learning. Foundation models (in particular LLMs) serve as the basis for fine-tuning on a new corpus of data. Since the original foundation models contain a very large number of parameters, fine-tuning them can be quite challenging. Development of the low-rank adaption technique (LoRA) for fine-tuning, and the quantized version of LoRA, also known as QLoRA, allows for fine-tuning of LLMs on a new smaller corpus of data. This paper focuses on the repeatability of fine-tuning four LLMs using QLoRA. We have fine-tuned them for seven trials each under the same hardware and software settings. We also validated our study for the repeatability (stability) issue by fine-tuning LLMs on two public datasets. For each trial, each LLM was fine-tuned on a subset of the dataset and tested on a holdout test set. Fine-tuning and inference were done on a single GPU. Our study shows that fine-tuning of LLMs with the QLoRA method is not repeatable (not stable), such that different fine-tuned runs result in different performance on the holdout test set.

The progress of incorporating deep learning in the field of medical image interpretation has been greatly hindered due to the tremendous cost and time associated with generating ground truth for supervised machine learning, alongside concerns about the inconsistent quality of images acquired. Active learning offers a potential solution to these problems of expanding dataset ground truth by algorithmically choosing the most informative samples for ground truth labeling. Still, this effort incurs the costs of human labeling, which needs minimization. Furthermore, automatic labeling approaches employing active learning often exhibit overfitting tendencies while selecting samples closely aligned with the training set distribution and excluding out-of-distribution samples, which could potentially improve the model’s effectiveness. We propose that the majority of out-of-distribution instances can be attributed to inconsistent cross images. Since the FDA approved the first whole-slide image system for medical diagnosis in 2017, whole-slide images have provided enriched critical information to advance the field of automated histopathology. Here, we exemplify the benefits of a novel deep learning strategy that utilizes high-resolution whole-slide microscopic images. We quantitatively assess and visually highlight the inconsistencies within the whole-slide image dataset employed in this study. Accordingly, we introduce a deep learning-based preprocessing algorithm designed to normalize unknown samples to the training set distribution, effectively mitigating the overfitting issue. Consequently, our approach significantly increases the amount of automatic region-of-interest ground truth labeling on high-resolution whole-slide images using active deep learning. We accept 92% of the automatic labels generated for our unlabeled data cohort, expanding the labeled dataset by 845%. Additionally, we demonstrate expert time savings of 96% relative to manual expert ground-truth labeling.

Large language models (LLM) are perceived to offer promising potentials for automating security tasks, such as those found in security operation centers (SOCs). As a first step towards evaluating this perceived potential, we investigate the use of LLMs in software pentesting, where the main task is to automatically identify software security vulnerabilities in source code. We hypothesize that an LLM-based AI agent can be improved over time for a specific security task as human operators interact with it. Such improvement can be made, as a first step, by engineering prompts fed to the LLM based on the responses produced, to include relevant contexts and structures so that the model provides more accurate results. Such engineering efforts become sustainable if the prompts that are engineered to produce better results on current tasks, also produce better results on future unknown tasks. To examine this hypothesis, we utilize the OWASP Benchmark Project 1.2 which contains 2,740 hand-crafted source code test cases containing various types of vulnerabilities. We divide the test cases into training and testing data, where we engineer the prompts based on the training data (only), and evaluate the final system on the testing data. We compare the AI agent's performance on the testing data against the performance of the agent without the prompt engineering. We also compare the AI agent's results against those from SonarQube, a widely used static code analyzer for security testing. We built and tested multiple versions of the AI agent using different off-the-shelf LLMs -- Google's Gemini-pro, as well as OpenAI's GPT-3.5-Turbo and GPT-4-Turbo (with both chat completion and assistant APIs). The results show that using LLMs is a viable approach to build an AI agent for software pentesting that can improve through repeated use and prompt engineering.

Histopathological classification in prostate cancer remains a challenge with high dependence on the expert practitioner. We develop a deep learning (DL) model to identify the most prominent Gleason pattern in a highly curated data cohort and validate it on an independent dataset. The histology images are partitioned in tiles (14,509) and are curated by an expert to identify individual glandular structures with assigned primary Gleason pattern grades. We use transfer learning and fine-tuning approaches to compare several deep neural network architectures that are trained on a corpus of camera images (ImageNet) and tuned with histology examples to be context appropriate for histopathological discrimination with small samples. In our study, the best DL network is able to discriminate cancer grade (GS3/4) from benign with an accuracy of 91%, F₁-score of 0.91 and AUC 0.96 in a baseline test (52 patients), while the cancer grade discrimination of the GS3 from GS4 had an accuracy of 68% and AUC of 0.71 (40 patients).