CT-based radiomics deep learning signatures for non-invasive prediction of metastatic potential in pheochromocytoma and paraganglioma: a multicohort study
Abstract Objectives This study aimed to develop and validate CT-based radiomics deep learning signatures for the non-invasive prediction of metastatic potential in pheochromocytomas and paragangliomas (PPGLs). Methods We conducted a retrospective analysis of 249 PPGL patients from three institutions...
Saved in:
| Main Authors: | , , , , , , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
SpringerOpen
2025-04-01
|
| Series: | Insights into Imaging |
| Subjects: | |
| Online Access: | https://doi.org/10.1186/s13244-025-01952-4 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Abstract Objectives This study aimed to develop and validate CT-based radiomics deep learning signatures for the non-invasive prediction of metastatic potential in pheochromocytomas and paragangliomas (PPGLs). Methods We conducted a retrospective analysis of 249 PPGL patients from three institutions, dividing them into training (n = 138), test1 (n = 71), and test2 (n = 40) sets. Based on the grading system for adrenal pheochromocytoma and paraganglioma (GAPP), patients were classified into low-risk (GAPP < 3) and high-risk (GAPP ≥ 3) groups. Radiomic features were extracted from CT venous phase images and modeled using six machine learning algorithms. The maximum 2D sections and 3D images of each tumor were input into four ResNet models to obtain predictive probabilities. Optimal models were selected based on receiver operating characteristic analysis and integrated with radiological features to develop a combined model, which was evaluated on external datasets, and explored prognostic information. Results The support vector machine radiomics and 2D ResNet-50 models demonstrated good performance. By integrating these two models with intratumoral necrosis features, we constructed a combined model that achieved high accuracy, with area under the curve (AUC) values of 0.90 for the training, 0.86 for the test1, and 0.88 for the test2 sets. This model effectively stratified patients based on metastasis-free survival (p = 0.003). Its predictive ability remains robust below the 6 cm threshold, with AUC values exceeding 0.87 across all datasets. Conclusions The combined model can predict the metastatic potential of PPGL in the preoperative stage, providing a precise surgical strategy for pheochromocytoma regarding the 6 cm surgical threshold. Critical relevance statement The combined model, established based on radiomic and deep learning signatures, shows potential for early preoperative prediction of metastatic potential in PPGL. Key Points Metastatic potential of PPGL affects surgical approaches and prognosis. CT-based radiomics deep learning signatures can predict the metastatic potential in PPGL.3. The combined model’s predictive ability remains robust below the 6-cm threshold. The combined model’s predictive ability remains robust below the 6-cm threshold. Graphical Abstract |
|---|---|
| ISSN: | 1869-4101 |