Deep learning-based cerebellar segmentation on T2-weighted fluid-attenuated inversion recovery magnetic resonance imaging for detecting cerebellar hypoplasia/atrophy in infants.
Authors
Affiliations (4)
Affiliations (4)
- Department of Radiology, Peking University First Hospital, No. 8 Xishiku Street, Beijing, 100034, China.
- School of Basic Medical Sciences, Capital Medical University, Beijing, China.
- Beijing Smart Tree Medical Technology Co., Ltd., Beijing, China.
- Department of Radiology, Peking University First Hospital, No. 8 Xishiku Street, Beijing, 100034, China. [email protected].
Abstract
Accurate measurement of cerebellar volume is crucial for diagnosing cerebellar atrophy or hypoplasia in infants. Although deep learning has achieved some success in cerebellar segmentation, existing studies primarily focus on adults. Challenges remain in applying these techniques to infants, especially those under 2 years old, due to the small cerebellar size and low tissue contrast in magnetic resonance imaging (MRI). As a result, developing segmentation models to identify abnormal cerebellar volumes in infants remains an unmet need. To develop and validate a deep learning model for cerebellar volume measurement in infant brain MRI across different ages, with a focus on detecting cerebellar hypoplasia or atrophy. A deep learning segmentation model was developed using a publicly available dataset of 558 neonatal MRI scans. The model was validated on two independent datasets: a normal set (492 scans of typical infant brains) and an abnormal set (40 scans of cerebellar hypoplasia or atrophy). Two radiologists manually refined the segmented cerebellar regions to ensure consistency, followed by quantification of cerebellar volumes and diameters. The model's segmentation accuracy was assessed using the Dice similarity coefficient (DSC). A linear regression model was developed using the normal set to predict subjects' ages based on actual age, sex, and cerebellar volume. The difference between the predicted and actual ages (Δage) was calculated and used to classify subjects as normal or abnormal. The diagnostic performance of Δage in distinguishing between the normal and abnormal sets was assessed using the area under the receiver operating characteristic (ROC) curve (AUC). The mean DSC for cerebellar segmentation was 0.962 in the normal set and 0.882 in the abnormal set. In the normal set, cerebellar diameters were as follows: the left-right diameter ranged from 44.688 mm to 102.266 mm, the anterior-posterior diameter from 32.656 mm to 69.180 mm, and the superior-inferior diameter from 24.000 mm to 61.000 mm. The average cerebellar volume in the normal set was 79.305 cm<sup>3</sup>, showing rapid growth from birth to 10.0 months of age, with a slower growth rate from 10.0 months to 24.0 months. The AUC for Δage in identifying subjects with cerebellar hypoplasia or atrophy was 0.851. The proposed deep learning model can accurately segment the cerebellum in infant brain MRI, quantify cerebellar volumes, and assist in identifying cerebellar hypoplasia or atrophy.