Dark-Field X-Ray Imaging Significantly Improves Deep-Learning based Detection of Synthetic Early-Stage Lung Tumors in Preclinical Models

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Dey, Joyoni, Meyer, Hunter C., Taqi, Murtuza S.
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866908622810775552
author Dey, Joyoni
Meyer, Hunter C.
Taqi, Murtuza S.
author_facet Dey, Joyoni
Meyer, Hunter C.
Taqi, Murtuza S.
contents Low-dose computed tomography (LDCT) is the current standard for lung cancer screening, yet its adoption and accessibility remain limited. Many regions lack LDCT infrastructure, and even among those screened, early-stage cancer detection often yield false positives, as shown in the National Lung Screening Trial (NLST) with a sensitivity of 93.8 percent and a false-positive rate of 26.6 percent. We aim to investigate whether X-ray dark-field imaging (DFI) radiograph, a technique sensitive to small-angle scatter from alveolar microstructure and less susceptible to organ shadowing, can significantly improve early-stage lung tumor detection when coupled with deep-learning segmentation. Using paired attenuation (ATTN) and DFI radiograph images of euthanized mouse lungs, we generated realistic synthetic tumors with irregular boundaries and intensity profiles consistent with physical lung contrast. A U-Net segmentation network was trained on small patches using either ATTN, DFI, or a combination of ATTN and DFI channels. Results show that the DFI-only model achieved a true-positive detection rate of 83.7 percent, compared with 51 percent for ATTN-only, while maintaining comparable specificity (90.5 versus 92.9 percent). The combined ATTN and DFI input achieved 79.6 percent sensitivity and 97.6 percent specificity. In conclusion, DFI substantially improves early-tumor detectability in comparison to standard attenuation radiography and shows potential as an accessible, low-cost, low-dose alternative for pre-clinical or limited-resource screening where LDCT is unavailable.
format Preprint
id arxiv_https___arxiv_org_abs_2510_27679
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Dark-Field X-Ray Imaging Significantly Improves Deep-Learning based Detection of Synthetic Early-Stage Lung Tumors in Preclinical Models
Dey, Joyoni
Meyer, Hunter C.
Taqi, Murtuza S.
Medical Physics
Computer Vision and Pattern Recognition
Machine Learning
Image and Video Processing
Optics
Low-dose computed tomography (LDCT) is the current standard for lung cancer screening, yet its adoption and accessibility remain limited. Many regions lack LDCT infrastructure, and even among those screened, early-stage cancer detection often yield false positives, as shown in the National Lung Screening Trial (NLST) with a sensitivity of 93.8 percent and a false-positive rate of 26.6 percent. We aim to investigate whether X-ray dark-field imaging (DFI) radiograph, a technique sensitive to small-angle scatter from alveolar microstructure and less susceptible to organ shadowing, can significantly improve early-stage lung tumor detection when coupled with deep-learning segmentation. Using paired attenuation (ATTN) and DFI radiograph images of euthanized mouse lungs, we generated realistic synthetic tumors with irregular boundaries and intensity profiles consistent with physical lung contrast. A U-Net segmentation network was trained on small patches using either ATTN, DFI, or a combination of ATTN and DFI channels. Results show that the DFI-only model achieved a true-positive detection rate of 83.7 percent, compared with 51 percent for ATTN-only, while maintaining comparable specificity (90.5 versus 92.9 percent). The combined ATTN and DFI input achieved 79.6 percent sensitivity and 97.6 percent specificity. In conclusion, DFI substantially improves early-tumor detectability in comparison to standard attenuation radiography and shows potential as an accessible, low-cost, low-dose alternative for pre-clinical or limited-resource screening where LDCT is unavailable.
title Dark-Field X-Ray Imaging Significantly Improves Deep-Learning based Detection of Synthetic Early-Stage Lung Tumors in Preclinical Models
topic Medical Physics
Computer Vision and Pattern Recognition
Machine Learning
Image and Video Processing
Optics
url https://arxiv.org/abs/2510.27679