MRI-based radiomics prediction of somatostatin receptor ligand response in acromegaly: an automated pipeline study.
Authors
Affiliations (12)
Affiliations (12)
- Endocrine Unit, Hospital de Vic, Vic 08500, Spain.
- Comparative Medicine and Bioimage Centre of Catalonia (CMCiB), Germans Trias Research Institute (IGTP), Badalona 08916, Spain.
- Biomedical Imaging Unit FISABIO-CIPF, Foundation for the Promotion of Research in Healthcare and Biomedicine (FISABIO), Valencia 46020, Spain.
- Service of Endocrinology and Nutrition, Ramón y Cajal University Hospital, Madrid 28034, Spain.
- Service of Endocrinology and Nutrition, Vall d'Hebrón University Hospital, Barcelona 08035, Spain.
- Neuroradiology Unit, Vall d'Hebrón University Hospital, Barcelona 08035, Spain.
- Service of Endocrinology and Nutrition, Germans Trias University Hospital, Badalona 08916, Spain.
- Service of Endocrinology, Hospital Central, Instituto de Previsión Social, Asunción 001531, Paraguay.
- Service of Endocrinology and Nutrition, Hospital La Princesa, Madrid 28006, Spain.
- Department of Radiology, Germans Trias University Hospital, Badalona 08916, Spain.
- Department of Medicine, Imaging Diagnostic Area, Universitat Internacional de Catalunya, Sant Cugat del Valles 08195, Spain.
- Department of Medicine, Universitat Autònoma de Barcelona, Campus de Can Ruti, Badalona 08916, Spain.
Abstract
Radiomics may capture tumor characteristics predicting therapeutic response. To develop an MRI-based radiomics pipeline for predicting response to SRLs in acromegaly. MRI data from 267 subjects across 3 datasets were used to develop an automated tumor detection and segmentation pipeline (Spanish multicenter acromegaly IGTP cohort, <i>n</i> = 81; 2 public datasets, <i>n</i> = 186). The 3-stage pipeline comprised: (1) automated tumor detection (YOLOv8), (2) 3D segmentation (SegResNet), and (3) radiomic feature extraction (PyRadiomics) with supervised classification. Treatment-response radiomics included contrast-enhanced T1-weighted (Ce-T1W1) MRI from 49 patients treated with SRL. SRL response was defined as ≥50% IGF-1 reduction or normalization. Eleven machine-learning classifiers were evaluated using stratified repeated <i>k</i>-fold cross-validation. Automated segmentation achieved a Dice coefficient of 0.795 using combined Ce-T1WI and T2-weighted images (T2WI), decreasing to 0.765 in the acromegaly-specific cohort. The best-performing model was a 3-feature logistic regression classifier, with an out-of-fold AUC of 0.798 (95% CI: 0.671-0.918), balanced accuracy of 0.793, <i>F</i>1-score of 0.815, and correct classification of 39/49 patients (79.6%). Performance improved as the radiomic feature set was reduced from 5 to 2. The most consistent feature was log-sigma-3-0-mm-3D_glcm_ClusterShade, which appeared in 98% of cross-validation folds and contributed most to the model-derived discriminative signal. In this multicenter radiomics study of SRL response prediction in acromegaly, an automated pipeline achieved good discriminative performance (AUC = 0.798). These findings highlight both the promise and the current challenges of radiomics-based treatment prediction.