LUCID: Lipid-featured Urine Cytology with Intelligent Detection for Non-Invasive Early Diagnosis of Bladder Cancer
- 1Beihang University
Abstract
Leveraging its non-invasive and convenient nature, urinary cytology offers a way for the early diagnosis and surveillance of bladder cancer (BCa). However, its sensitivity remains substantially inferior to the histopathological gold standard, rendering it inadequate to support precise clinical diagnosis. Here, we developed Lipid-featured Urine Cytology with Intelligent Detection (LUCID), an automated platform integrating single-cell stimulated Raman scattering (SRS) imaging and explainable AI model for accurate BCa detection. Following a newly developed standardized urine-preprocessing protocol, 19,678 urinary exfoliated cells from 26 BCa patients and 54 non-BCa participants were imaged using two-color SRS microscopy to simultaneously visualize lipid and protein distributions. For cell-level analysis, a customized deep-supervised Cellpose segmentation algorithm and a YOLO-based detection model were developed to enable precise single-cell segmentation and intracellular lipid droplet (LD) identification, respectively. Totally 20 quantitative morphological and compositional features per cell were extracted and 10 related principal components were obtained via principal component analysis. These features were subsequently integrated into a multiple-instance learning (MIL)-XGBoosting model, which achieved robust diagnostic performance, with 92.31% sensitivity, 87.04% specificity, and an AUC of 0.940. We found that morphological and compositional features provided complementary diagnostic information, while LD-associated features provided primary discriminative information, supporting intracellular LD phenotypes as potential urinary cellular markers of BCa. More importantly, LUCID demonstrated exceptional accuracy in detecting early-stage and small tumors, and showed higher concordance than FISH in a preliminary paired subset (5/6 vs. 2/6). Together, LUCID establishes a chemical imaging-AI paradigm that transforms molecular urine cytology.
Related
- Florbetaben PET for early and non-invasive diagnosis of light chain amyloidosis: a case report and brief review of the literature.
- A biomimetic AuNPs@MBene-TFT artificial tongue for sensitive non-invasive salivary microRNA-27a detection in HNSCC assessment
- Cellular, chromatin-accessibility, and spatial contextualization of bulk transcriptional patterns linked to positive peritoneal cytology in endometrial cancer
- Artificially intelligent modeling of macromolecular assemblies
- Electrochemical nucleic acid biosensors: Advancing blood-based diagnostics toward intelligent healthcare in resource-limited settings
BibTeX
@misc{hao2026lucid,
title = {{LUCID: Lipid-featured Urine Cytology with Intelligent Detection for Non-Invasive Early Diagnosis of Bladder Cancer}},
author = {Z. Hao and J. Liu and T. Fang and Y. Zheng and H. Hu and S. Hu and X. Zhang and L. Zhou and X. Chen and L. Yao and S. Yue},
howpublished = {medRxiv},
year = {2026},
month = oct,
doi = {10.64898/2026.10.14.26365504},
url = {https://www.medrxiv.org/content/10.64898/2026.10.14.26365504v1},
note = {CC BY-NC},
}
RIS
TY - UNPB
TI - LUCID: Lipid-featured Urine Cytology with Intelligent Detection for Non-Invasive Early Diagnosis of Bladder Cancer
AU - Z. Hao
AU - J. Liu
AU - T. Fang
AU - Y. Zheng
AU - H. Hu
AU - S. Hu
AU - X. Zhang
AU - L. Zhou
AU - X. Chen
AU - L. Yao
AU - S. Yue
PY - 2026
DA - 2026/10/14
PB - medRxiv
DO - 10.64898/2026.10.14.26365504
UR - https://www.medrxiv.org/content/10.64898/2026.10.14.26365504v1
AB - Leveraging its non-invasive and convenient nature, urinary cytology offers a way for the early diagnosis and surveillance of bladder cancer (BCa). However, its sensitivity remains substantially inferior to the histopathological gold standard, rendering it inadequate to support precise clinical diagnosis. Here, we developed Lipid-featured Urine Cytology with Intelligent Detection (LUCID), an automated platform integrating single-cell stimulated Raman scattering (SRS) imaging and explainable AI model for accurate BCa detection. Following a newly developed standardized urine-preprocessing protocol, 19,678 urinary exfoliated cells from 26 BCa patients and 54 non-BCa participants were imaged using two-color SRS microscopy to simultaneously visualize lipid and protein distributions. For cell-level analysis, a customized deep-supervised Cellpose segmentation algorithm and a YOLO-based detection model were developed to enable precise single-cell segmentation and intracellular lipid droplet (LD) identification, respectively. Totally 20 quantitative morphological and compositional features per cell were extracted and 10 related principal components were obtained via principal component analysis. These features were subsequently integrated into a multiple-instance learning (MIL)-XGBoosting model, which achieved robust diagnostic performance, with 92.31% sensitivity, 87.04% specificity, and an AUC of 0.940. We found that morphological and compositional features provided complementary diagnostic information, while LD-associated features provided primary discriminative information, supporting intracellular LD phenotypes as potential urinary cellular markers of BCa. More importantly, LUCID demonstrated exceptional accuracy in detecting early-stage and small tumors, and showed higher concordance than FISH in a preliminary paired subset (5/6 vs. 2/6). Together, LUCID establishes a chemical imaging-AI paradigm that transforms molecular urine cytology.
ER -