Development and Validation of a Combined MRI Radiomics, Imaging and Clinical Parameter-Based Machine Learning Model for Identifying Idiopathic Central Precocious Puberty in Girls

Abstract

Background: Idiopathic central precocious puberty (ICPP) impairs child development, without early intervention. The current reference standard, the gonadotropin-releasing hormone stimulation test, is invasive which may hinder diagnosis and intervention.

Purpose: To develop a model for accurate diagnosis of ICPP, by integrating pituitary MRI, carpal bone age, gonadal ultrasound, and basic clinical data.

Study type: Retrospective.

Population: A total of 492 girls with PP (185 with ICPP and 307 peripheral precocious puberty [PPP]) were randomly divided by reference standard into training (75%) and internal validation (25%) data. Fifty-one subjects (16 with ICPP, 35 with PPP) provided by another hospital as external validation.

Field strength/sequence: T1-weighted (spin echo [SE], fast SE, cube) and T2-weighted (fast SE-fat suppression) imaging at 3.0 T or 1.5 T.

Assessment: Radiomics features were extracted from pituitary MRI after manual segmentation. Carpal bone age, ovarian, follicle and uterine volumes and endometrium presence were assessed from radiographs and gonadal ultrasound. Four machine learning methods were developed: a pituitary MRI radiomics model, an integrated image model (with pituitary MRI, gonadal ultrasound and bone age), a basic clinical model (with age and sex hormone data), and an integrated multimodal model combining all features.

Statistical tests: Intraclass correlation coefficients were used to assess consistency of segmentation. Receiver operating characteristic (ROC) curves and the Delong tests were used to assess and compare the diagnostic performance of models. P < 0.05 was considered statistically significant.

Results: The area under of the ROC curve (AUC) of the pituitary MRI radiomics model, integrated image model, basic clinical model, and integrated multimodal model in the training data was 0.668, 0.809, 0.792, and 0.860. The integrated multimodal model had higher diagnostic efficacy (AUC of 0.862 and 0.866 for internal and external validation).

Conclusion: The integrated multimodal model may have potential as an alternative clinical approach to diagnose ICPP.

Evidence level: 3.

Technical efficacy: Stage 2.

Keywords: diagnostic model; machine learning; pituitary MRI; precocious puberty; radiomics.