A stacking model for AI-assisted diagnosis of suspected pituitary microadenomas on non-contrast T1COR MRI: a multicenter reader study on bridging the experience gap.
Authors
Affiliations (8)
Affiliations (8)
- Department of Magnetic Resonance, The Second Hospital & Clinical Medical School, Lanzhou University, Lanzhou, China.
- Gansu Province Clinical Research Center for Functional and Molecular Imaging, Lanzhou, China.
- Gansu Medical MRI Equipment Application Industry Technology Center, Lanzhou, China.
- Xiaogan Central Hospital, Xiaogan, China.
- Lincang People's Hospital, Lincang, China.
- Department of Magnetic Resonance, The Second Hospital & Clinical Medical School, Lanzhou University, Lanzhou, China. [email protected].
- Gansu Province Clinical Research Center for Functional and Molecular Imaging, Lanzhou, China. [email protected].
- Gansu Medical MRI Equipment Application Industry Technology Center, Lanzhou, China. [email protected].
Abstract
This study sought to quantify, through a multi-reader study, whether AI assistance improves diagnostic accuracy across experience levels, reduces bidirectional errors, and enhances inter-reader consensus in suspected pituitary microadenoma diagnosis. To this end, we developed and validated a stacking model integrating clinical, radiomics, and deep learning features on non-contrast T1COR MRI. This retrospective multicenter study enrolled 636 patients from three centers, divided into training (n = 321), internal validation (n = 138), and two external validation cohorts (n = 136, n = 41). We developed four base models-handcrafted radiomics, deep transfer learning (DTL), deep learning radiomics (DLR), and clinical-and integrated them via a stacking ensemble with logistic regression as the meta-classifier. To evaluate real-world clinical impact, a three-round reader study was conducted with 315 lesions and five radiologists (three juniors, two seniors). Readers assessed non-contrast T1COR MRI unaided, with DTL assistance, and with combined model assistance. Performance metrics included AUC, accuracy, sensitivity, specificity, and inter-reader consensus. The combined model outperformed all single-modality approaches across validation cohorts, achieving AUCs of 0.818 (EVC1) and 0.899 (EVC2) with balanced sensitivity and specificity (EVC2: 0.929 and 0.833, respectively). In the reader study with 315 lesions and five radiologists, AI assistance significantly improved diagnostic accuracy across all experience levels. With combined model assistance, gains were more pronounced: seniors achieved 75.4-79.0% accuracy in the internal validation cohort (P-adj < 0.05), and up to 85.4% in the most challenging external cohort (P-adj < 0.05)- an absolute improvement of approximately 30% over unaided performance. The combined model also enhanced diagnostic consensus, with total scores from five radiologists showing systematic improvement for both microadenoma and non-microadenoma groups. The findings of this study suggest that our combined model holds promise as an effective tool to assist radiologists in the diagnosis of suspected pituitary microadenoma, providing a foundation for future clinical translation.