arXiv subjects

Cynthia Lo

Publications and source records attributed to Cynthia Lo.

3 recordsLinked to original sources

arXiv2025

Comprehensive language-image pre-training for 3D medical image understanding

In the 3D medical image domain, vision-language pre-training is used to create vision-language encoders (VLEs) that can support radiologists by retrieving patients with similar abnormalities, predicting likelihoods of abnormality, or, with downstream adaptation, generating radiological reports. While the methodology holds promise, three challenges limit the capabilities of current 3D VLEs: data scarcity due to privacy concerns, high computational costs resulting from the volumetric nature of the images, and a domain shift between the long reports used for training and the short prompts used during inference for, e.g., zero-shot classification. As a consequence, natural-image VLE recipes do not directly transfer to 3D medical imaging. In this paper, we overcome these challenges by injecting additional supervision via a report generation objective and combining vision-language with vision-only pre-training, allowing us to leverage both image-only and paired image-text 3D datasets. Further, we propose a novel loss that addresses the domain shift between long reports and short textual prompts. Through these additional objectives, paired with best practices of the 3D medical imaging domain, we develop the Comprehensive Language-Image Pre-training (COLIPRI) encoder family. Our COLIPRI encoders achieve state-of-the-art performance in report generation, semantic segmentation, classification probing, and zero-shot classification. The model weights and inference code are freely available at https://huggingface.co/microsoft/colipri.

cs.CV↗

arXiv2017

Metabolite patterns reveal regulatory responses to genetic perturbations

Genetic and environmental perturbation experiments have been used to study microbes in a bid to gain insight into transcriptional regulation, adaptive evolution, and other cellular dynamics. These studies have potential in enabling rational strain design. Unfortunately, experimentally determined intracellular flux distribution are often inconsistent or incomparable due to different experimental conditions and methodologies. Computational strain design relies on constraint-based reconstruction and analysis (COBRA) techniques to predict the effect of gene knockouts such as flux balance analysis (FBA), regulatory on/off minimization(ROOM), minimization of metabolic adjustment (MOMA), relative optimality in metabolic networks (RELATCH). Most of these knock-out prediction methods are based on conserving inherent flux patterns (between wild type and mutant) that are thought to be representative of the cellular regulatory structure. However, it has been recently demonstrated that these methods show poor agreement with experimental data. To improve the fidelity of knockout predictions and subsequent computational strain design, we developed REMEP, a metabolite-centric method. We demonstrate the improved performance of REMEP by comparing the different methods on experimental knockout data of E. coli, and S. cerevisiae grown in batch cultures. REMEP retains most of the features of earlier algorithms but is much more accurate in capturing cellular responses to genetic perturbations. A primary reason for this is that REMEP relies on the assumption that cellular regulatory structure leaves a signature on metabolite patterns and not just flux patterns. REMEP will also prove useful in uncovering novel insights into cellular regulation and control.

q-bio.MN↗

arXiv2014

Enhanced Thermoelectric Efficiency via Orthogonal Electrical and Thermal Conductances in Phosphorene

Thermoelectric devices that utilize the Seebeck effect convert heat flow into electrical energy and are highly desirable for the development of portable, solid state, passively-powered electronic systems. The conversion efficiencies of such devices are quantified by the dimensionless thermoelectric figure of merit (ZT), which is proportional to the ratio of a device's electrical conductance to its thermal conductance. High ZT (>2) has been achieved in materials via all-scale hierarchical architecturing. This efficiency holds at high temperatures (700K~900K) but quickly diminishes at lower temperatures. In this paper, a recently-fabricated two-dimensional (2D) semiconductor called phosphorene (monolayer black phosphorus) is assessed for its thermoelectric capabilities. First-principles and model calculations reveal that phosphorene possesses spatially-anisotropic electrical and thermal conductances. The prominent electrical and thermal conducting directions are orthogonal to one another, enhancing the ratio of these conductances. As a result, ZT can reach 2.5 (the criterion for commercial deployment) along the armchair direction of phosphorene at T=500K and is greater than 1 even at room temperature given moderate doping (~2 x 10^16 m-2). Ultimately, phosphorene stands out as an environmentally sound thermoelectric material with unprecedented qualities: intrinsically, it is a mechanically flexible material that converts heat energy with high efficiency at low temperatures (~ 300K) - one whose performance does not require any sophisticated engineering techniques.

cond-mat.mtrl-sci↗