3082
MRI-based radiomics in pituitary adenomas: pre-treatment prediction of prolactin expression1China-Japan Union Hospital of Jilin University, Changchun, China
Synopsis
To investigate the potential of radiomics features based on MRI in predicting the prolactin expression of pituitary adenomas before treatment. We build the logistic model and validation, which can offer a noninvasive approach to predict the PRL expression of pituitary adenomas by the way of machine learning. It may be a reference in the diagnosis, treatment and prognosis evaluation for some pituitary adenoma subtypes.
PURPOSE
To retrospectively investigate the potential of radiomics features based on MRI in predicting the prolactin (PRL) expression of pituitary adenomas.METHOD AND MATERIALS
We enrolled 103 patients who pathologically confirmed pituitary adenomas (02/2013-07/2018) (50 patients with PRL positive expression and 53 patients with PRL negative expression). The 3087 radiomics features were extracted from T1-weighted images (T1WI), T2-weighted images (T2WI) and contrast-enhanced T1-weighted images (CE-T1WI). Each sequence images contained 1029 radiomics features. We used Variance Threshold, AVONA and the Least absolute shrinkage and selection operator (Lasso) methods to select features that were significantly associated with the prolactin expression. The 5-fold cross validation strategy and logistic regression model were used in the machine learning process. The performance of the models was explored with respect to the receiver operating characteristics (ROC) curve.Workflow of radiomics model building is showed in Fig 1.RESULTS
We compared the diagnostic efficiency of radiomics models based on T1WI, T2WI, CE-T1WI images, and series combinations. According to the prediction results of radiomics models, we found that T1&T2&CE-T1WI combined images have the stable and superior potential to predict the PRL expression of pituitary adenomas. The radiomics signature based on T1WI&T2WI&CE-T1WI images, which contained 19 selected features（Fig 2）, was significantly associated with PRL expression. In the training cohort, the radiomics model demonstrated a promising performance (Area Under Curve (AUC): 0.755, 95%CI, 0.664 to 0.846, 0.679 specificity, 0.720 sensitively, 0.699 accuracy)(Fig 3, A). And the performance of radiomics model was also verified in the validation cohort (AUC: 0.767 95%CI, 0.645 to 0.889, 0.656 specificity, 0.676 sensitively, 0.666 accuracy)(Fig 3, B).DISCUSSION
Our study is offering a non-invasive way to discriminate PRL expression of pituitary adenomas before treatment. We found the gender and volume are relevant with PRL expression, which is similar to the previous study1. We compared the diagnostic efficiency of radiomics models based on T1WI, T2WI, CE-T1WI images, and sequence combinations. According to the prediction results of radiomics models, we found that T1&T2&CE-T1WI combined images have the stable and superior potential to predict the PRL expression of pituitary adenomas. As the PRL positive express pituitary adenomas, lactotroph adenomas is the most common subtype and dopamine agonist is highly effectively in the treatment2-4. Although serological detection of prolactin is the most common and objective index in the diagnosis of lactotroph adenomas. But the serum level of prolactin can be influenced by some particular physiological situations, factitious factors and pathological situations5, for example, stress, renal failure or hypothyroidism etc6. Furthermore, according to the previous study7-9, quite a number of non-functioning pituitary adenomas patients presented with hyperprolactinemia, which caused by stalk effect10. Radiomics can be the potential biomarker to discriminate the PRL expression before surgery. Until now, we did not find any research about the connection between radiomics features and PRL expression.
Besides, according to the research, it has been illustrated that PRL positive pituitary adenomas have lower MVD than other subtypes11. MVD has been declared as a clinical parameter for predicting prognosis during tumor therapy12. Edward H also considered that MVD is also a parameter of the pituitary adenoma apoplexy and medical treatment13.
In the study, we noticed that radiomics model based on T1&T2&CE-T1 got a relatively superior prediction results with LR classifier. The contrast enhanced images provide a positive effect in the prediction of PRL expression. A little different from Zhang’s study14, our study showed that features extracted from contrast enhanced images may be helpful in the diagnosis of some specific subtypes of pituitary adenomas like PRL positive pituitary adenomas. That means some radiomics features of contrast enhanced features may be correlated with some particular biomarkers of pituitary adenomas. Our study also had some limitations. The number of enrolled patients is limited and our study is a single-center research. Multiple-center approach is crucial to obtain a sufficient sample size and reduce the regional difference.
CONCLUSION
Our study provides a promising approach to predict PRL expression before treatment, which may assist the decision-making of clinical therapy and follow-up.Acknowledgements
The authors gratefully acknowledge the colleagues who accomplish the MR scan in China-Japan Union Hospital of Jilin University. We would like to express our thanks to the technical support of Huiying Medical Technology Co. Ltd.References
1. Huang Y, Ding C, Zhang F, et al. Role of prolactin/adenoma maximum diameter and prolactin/adenoma volume in the differential diagnosis of prolactinomas and other types of pituitary adenomas[J]. Oncology Letters, 2018, 15(2):2010.
2. Melmed S. PITUITARY TUMORS[J]. Endocrinology & Metabolism Clinics of North America, 2015, 44(1):1-9.
3. Rogers A, Karavitaki N, Wass J A H. Diagnosis and management of prolactinomas and non-functioning pituitary adenomas[J]. Bmj, 2014, 349(sep10 7):g5390.
4. Melmed S, Casanueva F F, Hoffman A R, et al. Diagnosis and treatment of hyperprolactinemia: an Endocrine Society clinical practice guideline.[J]. Journal of Clinical Endocrinology & Metabolism, 2011, 96(2):273.
5. Saleem M, Martin H, Coates P, Prolactin Biology and Laboratory Measurement: An Update on Physiology and Current Analytical Issues, The Clinical biochemist. Reviews 39(1) (2018) 3-16.
6. Capozzi A, Scambia G, Pontecorvi A, et al. Hyperprolactinemia: pathophysiology and therapeutic approach.[J]. Gynecological Endocrinology, 2015, 31(7):506-510.
7. Zhang F, Huang Y, Ding C, et al. The prevalence of hyperprolactinemia in non-functioning pituitary macroadenomas.[J]. Int J Clin Exp Med, 2015, 8(10):18990-18997.
8. Hong J W, Mi K L, Sun H K, et al. Discrimination of prolactinoma from hyperprolactinemic non-functioning adenoma[J]. Endocrine, 2010, 37(1):140-147.
9. Behan L A, O'Sullivan E P, Glynn N, et al. Serum prolactin concentration at presentation of non-functioning pituitary macroadenomas.[J]. Journal of Endocrinological Investigation, 2013, 36(7):508-514..
10. Bergsneider M, Mirsadraei L, Yong W H, et al. The pituitary stalk effect: is it a passing phenomenon?[J]. J Neurooncol, 2014, 117(3):477-484.
11. Bălinişteanu B, Cimpean A M, Melnic E, et al. Crosstalk between tumor blood vessels heterogeneity and hormonal profile of pituitary adenomas: evidence and controversies.[J]. Anticancer Research, 2014, 34(10):5413.
12. Lan J, Li J, Ju X, et al. Relationship between microvessel density and cancer stem cells in tumor angiogenesis: a meta-analysis.[J]. Biomarkers in Medicine, 2016, 10(8):919-932.
13. Oldfield E H, Merrill M J. Apoplexy of pituitary adenomas: the perfect storm.[J]. Journal of Neurosurgery, 2015, 122(6):1444-9.
14. Zhang S, Song G, Zang Y, et al. Non-invasive radiomics approach potentially predicts non-functioning pituitary adenomas subtypes before surgery[J]. European Radiology, 2018:1-10.
Figures
