Endocrine Abstracts

The introduction of the brain magnetic resonance imaging (MRI) of into routine clinical practice has led to a significant increase of pituitary masses detection, including as incidental findings. An increase in the frequency of detection of pituitary gland formations entails a significant increase in costs for clinical and hormonal examination, MRI monitoring and dynamic observation by an endocrinologist. However, only a small number of identified formations require radical treatment. An important clinical problem is to determine the malignancy, predictors of potential growth and hormonal activity of pituitary incidentalomas. The development of an automated algorithm for diagnosing pituitary formations to identify groups of patients for priority examination is of great practical importance.

Materials and methods: A cross-sectional study of 746 patients was conducted for development a neural network program for automated MRI pituitary formations screening. All patients underwent the pituitary MRI without contrast at the Republican Clinical Hospital of Medical Rehabilitation in 2019-2022, coronal (T1, T2) and sagittal (T1) projections were studied. The MRI assessment was carried out by a radiologist; the clinical diagnosis was verified by endocrinologists.16 000 (30%) MR images from the original sample were manually labeled. Neural network architecture has been developed that is capable of performing multi-class classification using segmented images of the pituitary gland obtained as the output of the Faster-RCNN neural network (pretrained model) with a ResNet 50-FPN framework. The ratio of the intersection area of the predicted area and that obtained as a result of manual marking to their merging area, averaged over all images, was used as a metric. The test sample size was 25% of the original one.

Results: The average age 40.9±16.5 years. Structure of diagnosis: 176 – normal, 336 – microadenoma, 68 – macroadenoma, 108 – postoperative changes, 58 – other pathology of the pituitary gland. A trained neural network analyzed MRI of pituitary formations.

Results: Accuracy (proportion of correct answers) by class: normal – 84%, microadenoma – 77%, microadenoma – 90%, postoperative changes – 91%, others – 91% (total – 86%). Precision (accuracy)=0.63, Recall (completeness)=0.7, F1-measure=0.66. Thus, the efficiency of predictions is sufficient for automated screening of the proposed groups of pituitary formations.

Conclusions: The results of the study prove the possibility of MRI pituitary formations automated screening using a developed neural network with a high degree of reliability.