Automatic Fetal Brain Landmark Detection for Gestational Age Prediction

Abstract

During gestation, the human fetal brain undergoes dynamic transformations like an increase in size, changes in shape due to cortical folding processes, or progression of myelination. Recent research has utilized magnetic resonance imaging (MRI) to capture and analyze these processes.

Until now, the age of fetuses, necessary to correctly classify the stage of development, had to be determined manually by experts, via inspection of the individual scans and determining different anatomical landmarks. The relative positions of these landmarks enable the automatic prediction of fetal age, especially with modern machine learning techniques. [1]

However, the majority of available brain age prediction frameworks are concerned with the anatomy of the adult brain and degenerative disease thereof and are thus not directly applicable to the developing fetal brain. [2], [3]

This project aims to develop a framework for automatic landmark detection based on fetal brain MRI scans for automatic fetal age classification. 

The proposed approach consists of two components: a landmark prediction component and an age classification component. The former predicts five landmarks in the fetal brain, namely the left and right anterior horn, the left and right posterior horn as well as the cerebellum. For the landmark prediction, different models are tested. Support Vector Machines and Random Forests are used as baseline supervised learning algorithms, and a state-of-the-art deep learning approach involving Convolutional Neural Networks is used to compare different approaches to the problem. The landmark detection algorithm is evaluated by calculating the Euclidean distance between the predicted landmarks and target landmarks and further improved to minimize the distance. Different setups with varying hyperparameter combinations and data augmentation techniques, e.g. rotating or cropping the images, are investigated for improving model performance.

Based on the predicted landmarks, the age classification component is trained to classify the fetal age. This model builds upon existing frameworks found in the literature, for example supervised regression models [2], and extends these by adapting them to fetal brain MR images. The input of the age classification model are the predicted landmark coordinates. The output is the age in gestation weeks and days as a float value. The age classification is evaluated using the mean absolute error between the predicted age and the target age.

By automating the landmark prediction and age classification processes, annotation costs and personal efforts can be reduced. Additionally, an increase in research speed as well as an improvement in research quality are expected. The overall aim is to create an accessible and widely applicable framework for fetal age prediction.

References

[1] B. M. Kline-Fath, D. I. Bulas, and W. Lee, Fundamental and Advanced Fetal Imaging: Ultrasound and MRI. Philadelphia: Wolters Kluwer Health /Lippincott Williams & Wilkins, 2021.

[2] J. Han, S. Y. Kim, J. Lee, and W. H. Lee, ‘Brain Age Prediction: A Comparison between Machine Learning Models Using Brain Morphometric Data’, Sensors, vol. 22, no. 20, 2022. 

[3] L. Shen et al., ‘Attention-guided deep learning for gestational age prediction using fetal brain MRI’, Scientific Reports, vol. 12, no. 1408, 2022.