Fetal Brain MRI Measurements Using a Deep Learning Landmark Network with Reliability Estimation

Backgroud: Clinical assessment of fetal brain development based on MRI is mainly subjective and based on a few biometric linear measurements. Obtaining the measurements manually is time-consuming and is subject to observer variability. Our goal was to develop a fully automatic method to compute landmark-based linear measurements in a fetal brain MRI scan and to estimate their reliability.

Methods: We have developed a new deep learning method, FMLNet, that automatically computes linear measurements in a fetal brain MRI volume. The method is based on landmark detection and estimates their location reliability. It consists of four steps: 1) fetal brain region of interest detection with a two-stage anisotropic U-Net; 2) reference slice selection with a convolutional neural network (CNN); 3) linear measurement computation with a slice-wise landmarks detection with a CNN; 4) measurement reliability estimation using a Gaussian Mixture Model. The advantages of our method are that it does not rely on heuristics to identify the land-marks, that it does not require fetal brain structures segmentation, and that it is robust since it incorporates reliability estimation. We demonstrate our method on three key fetal biometric measurements from fetal brain MRI volumes: Cerebral Biparietal Diameter (CBD), Bone Bipari-etal Diameter (BBD), and Trans Cerebellum Diameter (TCD).

Results: Experimental results on training and test datasets of 164 and 50 fetal MRI volumes on which measurements were performed by an expert radiolo-gist yield a 95% confidence interval agreement of 2.25mm, 2.09mm and 2.17mm for CBD, BBD and TCD, all below the inter-observer variability.

Conclusion: Our method is generic, as it can be directly applied to other linear measurements in volumetric scans and can be used in a clinical setup.