Back to all papers

Dual-energy x-ray imaging for AI-based lung tumor localization using foundation models: a phantom study.

July 9, 2026pubmed logopapers

Authors

Alqethami N,Xie W,Blöcker TJ,Steininger P,Kurz C,Landry G,Palaniappan P,Riboldi M

Affiliations (5)

  • Faculty of Physics, Chair of Experimental Physics - Medical Physics, Ludwig-Maximilians-Universität München, Am Coulombwall 1, Garching b. München, Bavaria, 85748, Germany.
  • Radiation oncology, University Hospital Munich Campus Grosshadern, Marchioninistr. 15, Munchen, 81377, Germany.
  • medPhoton GmbH, Karolingerstraße 16, Salzburg, 5020, Austria.
  • Radiation Oncology, University Hospital Munich Campus Grosshadern, Marchioninistr. 15, Munchen, 81377, Germany.
  • Faculty of Physics, Chair of Experimental Physics - Medical Physics, Ludwig-Maximilians-Universitat Munchen, Am Coulombwall 1, Garching b. München, Bavaria, 85748, Germany.

Abstract

This study assessed the performance of AI-based tumor tracking using foundation models in planar dual-energy (DE) X-ray compared with the corresponding high-energy(HE) images from the same DE acquisition.&#xD;&#xD;Approach: An in-house phantom with pork ribs and artificial tumors was scanned using&#xD;fast kVp-switching DE planar imaging using two kVp pairs (60/120 kVp and 80/120 kVp),&#xD;with two tube current settings (15 mA and 20 mA) applied to both DE protocols.&#xD;Weighted logarithmic subtraction was applied to generate soft-tissue-enhanced DE images&#xD;by evaluating weighting factors in the range of 0.55-0.85. The contrast-to-noise ratio&#xD;(CNR) between bone and background regions was calculated to determine the optimal&#xD;weighting factor for bone suppression. An AI foundation model was then evaluated for&#xD;frame-by-frame tumor localization using first-frame initialization, and its predictions were&#xD;compared with manually generated ground-truth (GT) segmentation.&#xD;&#xD;Main result: The optimal weighting factor ranged from 0.57 to 0.79. Increasing the tube&#xD;current improved DE CNR, with median improvements of 4.7% for the 60/120 pair kVp&#xD;and 11.5% for the 80/120 kVp pair. The AI model achieved high localization accuracy&#xD;across DE and HE images (Dice scores > 0.9, HD95 < 2-3 mm). Although global metrics&#xD;showed limited differences between the two image types and across imaging protocols,&#xD;analysis of the missed tumor fraction, defined as the proportion of the GT tumor area not&#xD;captured by the predicted segmentation, demonstrated improved localization with DE&#xD;imaging, particularly for the smallest spherical tumor. For the 60/120 and 80/120 kVp&#xD;pairs, the median missed tumor fraction was 35-48% and 50-55% for HE images,&#xD;compared with 5-10% and <15% for DE images, respectively.&#xD;&#xD;Significance: This study demonstrates that integrating DE imaging with AI foundation models may improve tumor localization in anatomically challenging regions, particularly when tumor-bone overlap occurs.

Topics

Journal Article

Ready to Sharpen Your Edge?

Subscribe to join 11k+ peers who rely on RadAI Slice. Get the essential weekly briefing that empowers you to navigate the future of radiology.

We respect your privacy. Unsubscribe at any time.