[Paper Review] Deep BV: A Fully Automated System for Brain Ventricle Localization and Segmentation in 3D Ultrasound Images of Embryonic Mice

Ziming Qiu, Jack Langerman+6 more|arXiv (Cornell University)|Nov 5, 2018

Fetal and Pediatric Neurological DisordersMedicine22 references3 citations

TL;DR

This paper presents Deep BV, a fully automated deep learning system for localizing and segmenting brain ventricles (BV) in 3D high-frequency ultrasound (HFU) images of embryonic mice. It uses a two-stage approach: first, a 3D convolutional neural network (3D-CNN) localizes the BV within a volume via sliding window inference, followed by a fully convolutional network (FCN) for precise segmentation, achieving a Dice Similarity Coefficient (DSC) of 0.8956 on an unseen test set—25% higher than prior state-of-the-art.

ABSTRACT

Volumetric analysis of brain ventricle (BV) structure is a key tool in the study of central nervous system development in embryonic mice. High-frequency ultrasound (HFU) is the only non-invasive, real-time modality available for rapid volumetric imaging of embryos in utero. However, manual segmentation of the BV from HFU volumes is tedious, time-consuming, and requires specialized expertise. In this paper, we propose a novel deep learning based BV segmentation system for whole-body HFU images of mouse embryos. Our fully automated system consists of two modules: localization and segmentation. It first applies a volumetric convolutional neural network on a 3D sliding window over the entire volume to identify a 3D bounding box containing the entire BV. It then employs a fully convolutional network to segment the detected bounding box into BV and background. The system achieves a Dice Similarity Coefficient (DSC) of 0.8956 for BV segmentation on an unseen 111 HFU volume test set surpassing the previous state-of-the-art method (DSC of 0.7119) by a margin of 25%.

Motivation & Objective

• To address the challenge of manual, time-consuming, and expertise-dependent segmentation of brain ventricles in 3D high-frequency ultrasound (HFU) images of embryonic mice.
• To develop a fully automated system that eliminates human intervention in BV localization and segmentation for in utero embryonic imaging.
• To improve segmentation accuracy beyond existing state-of-the-art methods in the context of 3D HFU volumes.
• To enable high-throughput, reproducible volumetric analysis of brain ventricle development in preclinical neurodevelopmental studies.

Proposed method

• A 3D volumetric convolutional neural network (3D-CNN) is applied across a sliding window over the entire 3D HFU volume to detect a 3D bounding box enclosing the brain ventricle.
• The detected bounding box is cropped and fed into a fully convolutional network (FCN) for pixel-wise segmentation of BV from background.
• The system is trained end-to-end using annotated 3D HFU volumes of embryonic mice, with loss functions optimized for segmentation accuracy.
• The two-stage pipeline ensures computational efficiency by limiting high-resolution segmentation to a localized region of interest.
• The method leverages spatial context and hierarchical feature learning through 3D convolutions to capture complex anatomical structures in low-contrast ultrasound data.
• The architecture is designed to be robust to intensity variations and noise typical in high-frequency ultrasound imaging.

Experimental results

Research questions

• RQ1 Can a fully automated deep learning system achieve accurate brain ventricle localization and segmentation in 3D high-frequency ultrasound images of embryonic mice?
• RQ2 How does the performance of the proposed two-stage 3D-CNN and FCN pipeline compare to existing state-of-the-art methods in BV segmentation?
• RQ3 To what extent does the system reduce inter-rater variability and manual effort compared to traditional segmentation approaches?
• RQ4 Can the system generalize to unseen 3D HFU volumes without fine-tuning?
• RQ5 What is the impact of using a sliding window approach for 3D localization on segmentation accuracy and computational efficiency?

Key findings

• The proposed Deep BV system achieved a Dice Similarity Coefficient (DSC) of 0.8956 on an unseen test set of 111 3D HFU volumes, significantly outperforming the previous state-of-the-art method, which achieved a DSC of 0.7119.
• The improvement in DSC represents a 25% relative increase in segmentation accuracy, demonstrating the effectiveness of the two-stage 3D-CNN and FCN architecture.
• The system successfully localized and segmented brain ventricles in all 111 test volumes without manual intervention, confirming its robustness and automation capability.
• The use of a 3D sliding window for localization reduced computational overhead by focusing high-resolution segmentation on a relevant region of interest.
• The method demonstrated strong generalization to unseen data, indicating its potential for use in large-scale developmental neuroimaging studies.
• The results validate the utility of deep learning in overcoming challenges posed by low-contrast, noisy 3D ultrasound data in embryonic imaging.