A Convolutional Neural Network Model for the Prediction of Prolactinoma Response to Dopamine Agonists

Introduction: Prolactinomas, constituting the most prevalent type of pituitary adenomas, pose a significant clinical challenge due to their propensity for growth and the resulting disruption of hormonal balance. The primary therapeutic approach for managing prolactinomas involves the use of dopamine agonists (DAs), where surgical resection is considered a primary alternative, particularly for large tumors or when the tumor proves resistant to DA. However, a consensus regarding the optimal initial management strategy has yet to be established. This study's principal objective is to explore predictors of resistance to DA therapy, with the aim of assisting clinicians in making informed decisions regarding the initial management of prolactinomas.

Methods: We conducted a retrospective review of medical records from Emory University Hospital, identifying patients diagnosed with prolactinomas between January 2000 and December 2022. A review of eligible patients’ medical records was performed to extract demographic information, clinical presentations, radiological imaging, hormonal profiles, treatment records, and up to 5 years of follow-up data. The primary outcome measure was resistance to DA therapy, defined as persistent hyperprolactinemia and/or tumor growth despite optimal medical management. A convolutional neural network (CNN) models was constructed and trained on pretreatment DICOM images, using a split of 80 patients for training and 20 for validation. The model design, training, and validation were conducted using a curated Python code.

Results: The study cohort comprised 58 patients, with an average age of 42.6 ± 16.1 years, with 60% being females. Before initiating therapy, the median prolactin level was 683 ng/mL (interquartile range [IQR]: 324–1,354), and the median tumor volume was 5.2 cm³ (IQR: 1.5–10.6). Within this cohort, 24% of patients with DA therapy showed no tumor volume control and/or normalization of prolactin levels, despite compliance to therapy. Univariate logistic regression was employed to investigate predictors of failed DA therapy. Patient demographics, including age, gender, race, and comorbidities, displayed no significant correlation with DA failure (p > 0.05). Neither tumor volume nor the presence of optic chiasm or optic nerve compression on radiological imaging correlated with the odds of DA failure (p = 0.79 and p = 0.39, respectively). Lastly, the prolactin level before initiating therapy did not correlate with DA failure (p = 0.843). The CNN model showed high predictive accuracy in the prediction of DA response, especially when analyzing images from the coronal plane, with a root mean square error (RMSE) of 1.578 and a mean absolute error (MAE) of 2.768. The ROC area-under-curve was 0.8 and the coefficient of determination (R ²) was 0.63 (p < 0.05).

Conclusion: DA therapy in prolactinomas carries a risk of failure. Our analysis revealed no correlation between clinical patient characteristics and the failure of DA therapy. However, pretherapeutic radiological analysis using a CNN model demonstrated promise in predicting therapy response, aiding in pretreatment assessment and patient counseling. Larger multicenter studies are needed to externally validate these findings.

Publikationsverlauf

Artikel online veröffentlicht:
07. Februar 2025