The feasibility of Ki parametric imaging in clinical dynamic 18F‐FDG total‐body PET using a simulated‐data‐driven machine learning algorithm
Background Despite Ki parametric images providing high sensitivity and specificity in clinical diagnosis and therapeutic evaluation, their clinical application is limited by prolonged scan durations. Previous research has employed machine learning algorithms to generate reliable Ki images from 18F‐F...
Uloženo v:
| Vydáno v: | Medical physics (Lancaster) Ročník 52; číslo 10 |
|---|---|
| Hlavní autoři: | , , , , , , , , , , , , |
| Médium: | Journal Article |
| Jazyk: | angličtina |
| Vydáno: |
01.10.2025
|
| Témata: | |
| ISSN: | 0094-2405, 2473-4209, 2473-4209 |
| On-line přístup: | Získat plný text |
| Tagy: |
Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
|
| Shrnutí: | Background
Despite Ki parametric images providing high sensitivity and specificity in clinical diagnosis and therapeutic evaluation, their clinical application is limited by prolonged scan durations. Previous research has employed machine learning algorithms to generate reliable Ki images from 18F‐FDG PET scans with a duration of 10 min, but these algorithms require extensive real‐world data for effective training.
Purpose
This study explored the feasibility of a simulated‐data‐driven approach for clinical Ki parametric imaging with reduced scan duration.
Methods
A cluster analysis was conducted using K‐Means on 60‐minute dynamic total‐body PET data from 25 subjects, followed by the construction of a noise‐free simulated dataset using the Patlak equation. Subsequently, noise at varying levels was incorporated into the noise‐free simulated dataset, resulting in four distinct training sets. Each dataset contains 1 200 000 simulated Ki values as labels, along with corresponding tissue‐to‐blood concentration ratios categorized by noise level (free, low, middle, and high). Based on the same simulation methodology, we also generated four corresponding test datasets at different noise levels (each with identical size to the training dataset). Ultimately, XGBoost models, each trained with data at different noise levels, were employed to predict Ki values from short‐duration (50–60 min) PET scans. The accuracy of the Ki values generated from short‐duration scans by our proposed method was evaluated using both simulated data (the four test datasets at varying noise levels) and real‐world data (dynamic total‐body PET data from 25 subjects). The Ki images generated by the conventional Patlak method with a t* of 20 min post‐injection were chosen as the gold standard.
Results
Evaluation based on both simulated and real‐world data indicates that training algorithms with a noise‐inclusive simulated dataset significantly improve the average accuracy of Ki values. In evaluations with real‐world data, Ki images from 50–60‐minute dynamic PET scans generated using our proposed approach demonstrated superior performance compared to the conventional Patlak method, achieving a Pearson's correlation coefficient of 0.94 (vs. 0.42 for Patlak), a lower normalized mean square error of 0.11 (vs. 5.33), and a higher peak signal‐to‐noise ratio of 64.32 (vs. 47.87).
Conclusions
The simulated‐data‐driven approach can generate reliable Ki images from clinical 18F‐FDG dynamic total‐body PET scans, thereby reducing the costs associated with collecting real‐world data. |
|---|---|
| Bibliografie: | Wenjian Gu and Weiping Liu contributed equally to this article. ObjectType-Article-1 SourceType-Scholarly Journals-1 ObjectType-Feature-2 content type line 23 |
| ISSN: | 0094-2405 2473-4209 2473-4209 |
| DOI: | 10.1002/mp.70053 |