Abstract
Volumetric medical image segmentation, particularly in 3D ultrasound, is often hindered by low signal-to-noise ratios and high computational demands. We present a novel "segmentation-by-detection" framework that employs a two-stage cascade architecture. A 2D Region Proposal Network (RPN) first identifies regions of interest to create a 3D attention mask. This mask guides a 3D U-Net to perform precise segmentation by focusing on relevant anatomical structures. Our approach demonstrates superior performance in segmenting the infant femoral head, achieving an 8.7% improvement in Intersection over Union (IoU) compared to standard 3D U-Net architectures.
Introduction
Processing volumetric medical data (MRI, CT, 3D ultrasound) presents a significant challenge for deep learning models. Standard 2D slice-by-slice analysis often neglects vital inter-slice anatomical context, while full 3D volumetric approaches are memory-intensive and prone to poor convergence due to the high volume of irrelevant "noise" surrounding target tissues. This paper addresses these issues by introducing a spatial attention mechanism that focuses the network's resources on the most relevant regions of the volume.
Methodology
The proposed framework is a cascade network consisting of two primary modules: a 2D Detection Module and a 3D Segmentation Module.
- Detection Module (2D): Utilizing a Region Proposal Network (RPN) with a VGG-16 backbone, this module processes 2D slices to identify candidate bounding boxes (Regions of Interest, or RoIs).
- Attention Model: The 2D detections are stacked and concatenated to form a 3D binary mask, serving as a spatial attention map.
- Segmentation Module (3D): A 3D U-Net receives both the original 3D volume and the 3D attention mask as inputs. The mask acts as an additional channel, allowing the network to "distill" information from the identified region while ignoring peripheral noise.
This hybrid pipeline leverages the strengths of 2D networks (ease of training, pre-trained weights) to simplify the complex 3D segmentation task.
Experimental Setup
The model was evaluated on the clinical task of segmenting the femoral head in 3D ultrasound (3DUS) scans of infants, a key step in diagnosing Developmental Dysplasia of the Hip (DDH). The dataset comprised 19 3DUS scans from infants aged 4 to 111 days. Volumes were down-sampled to $128 \times 184 \times 96$. Ten scans were used for training, and nine were reserved for testing.
Results and Discussion
The "Segmentation-by-Detection" method significantly outperformed the baseline 3D U-Net across multiple metrics. The proposed method achieved an Average IoU of 0.709, compared to 0.622 for the standard 3D U-Net. Furthermore, the cascade network exhibited greater robustness in worst-case scenarios and demonstrated superior convergence properties.
The use of an attention mask effectively reduces the impact of low signal-to-noise ratios common in ultrasound imaging. By solving the localization problem first, the network can more efficiently learn the precise boundaries of complex anatomical structures.
Conclusion
The proposed cascade architecture effectively bridges the gap between 2D detection and 3D segmentation. While the current study is limited by a small dataset size, the results indicate that "segmentation-by-detection" is a promising approach for volumetric medical imaging. Future work may explore human-in-the-loop integration, where manual coarse localization could further enhance the precision of the 3D segmentation module in clinical environments.
Summary updated