💻 computer science

Dual-encoder cross-scale dependency fusion network for placenta segmentation using MR and ultrasound images

This paper proposes DECSDF-Net, a dual-encoder network incorporating specialized attention and fusion modules to effectively address challenges like ambiguous boundaries and modality differences, achieving state-of-the-art placenta segmentation performance on both MRI and ultrasound datasets.

Original authors: Sisi Song, Yanfen Xin, Hong Fan

Published 2026-09-30

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Sisi Song, Yanfen Xin, Hong Fan

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The placenta is a temporary organ that forms during pregnancy, acting as a vital lifeline between a mother and her developing baby. It handles the exchange of nutrients and oxygen while filtering out waste, and its health is a primary indicator of whether a pregnancy is progressing safely. When the placenta is not functioning correctly or is positioned in a way that blocks the birth canal, it can lead to serious complications for both mother and child. To understand the condition of this organ, doctors rely heavily on medical imaging, primarily using two different technologies: ultrasound, which uses sound waves to create pictures, and magnetic resonance imaging, which uses powerful magnets and radio waves. While these images are invaluable, the placenta itself is a difficult subject to study. It does not have a fixed shape, its texture varies from person to person, and the images often suffer from blurry edges or confusing noise that makes it hard to tell exactly where the organ begins and ends.

For decades, doctors have manually traced the outline of the placenta on these images to measure its size and location, a process that is slow, tiring, and prone to human error. In recent years, computers have begun to take on this task using artificial intelligence, specifically a type of software known as a neural network that learns to recognize patterns. However, teaching a computer to see the placenta is like trying to find a specific cloud in a stormy sky; the boundaries are often indistinct, and the organ looks different depending on whether it is viewed through sound waves or magnetic fields. Existing computer programs often struggle with these variations, sometimes missing parts of the organ or drawing lines in the wrong places. This is where a new study from researchers at Deyang People's Hospital and Sichuan Polytechnic University offers a significant step forward. They have developed a specialized computer system designed to automatically and accurately map the placenta in both ultrasound and magnetic resonance images, overcoming the visual confusion that has long plagued automated analysis.

The researchers built a system they call DECSDF-Net, which operates on a simple but powerful idea: using two different "eyes" to look at the same image at the same time. Instead of relying on a single path to process the picture, the system splits the image into two parallel streams. One stream focuses on capturing the broad, general shape of the placenta, while the other pays close attention to the fine, detailed edges. By running these two streams side by side, the computer can learn from the strengths of each approach, combining the big picture with the small details to form a complete understanding of what it is seeing. This dual approach allows the system to handle the messy, variable nature of placental images much better than previous methods that used only a single line of analysis.

To make this dual vision even sharper, the researchers added several specialized tools into the system's brain. One of these tools acts like a compass for the image, helping the computer understand the direction of the placenta's edges, whether they run horizontally or vertically. This is crucial because the placenta often has curved or irregular boundaries that standard software might miss. Another tool helps the system connect the dots between different layers of information. It takes the deep, abstract understanding of what the organ is and uses it to refine the shallow, pixel-level details, ensuring that the final outline is not just a guess but a precise match to the actual tissue. The system also includes a mechanism that looks at the image through different "lenses" simultaneously, allowing it to see the placenta as a whole while also zooming in on specific textures. Finally, as the computer builds its final map, it gradually combines all these different views, layering the information until a clear, complete picture emerges.

The team tested this new system on two large collections of real patient images: one set containing magnetic resonance scans and another containing ultrasound scans. They compared the computer's work against the gold standard, which is the manual tracing done by experienced human doctors. The results were remarkably consistent across both types of imaging. On the magnetic resonance images, the system correctly identified the placenta's location and shape with a high degree of overlap with the human experts, achieving a score of 0.8838 on a scale where 1.0 would be a perfect match. On the more difficult ultrasound images, which are often grainier and harder to read, the system still performed exceptionally well, scoring 0.8778. These numbers indicate that the system is not just guessing; it is reliably finding the organ in images where the boundaries are often unclear to the human eye.

To understand exactly which parts of their new system were responsible for this success, the researchers ran a series of tests where they removed one tool at a time. They found that without the dual-encoder design, the system struggled to capture both the shape and the details. When they removed the tool that helped connect different layers of information, the accuracy dropped noticeably. Each component they added back improved the final result, proving that the combination of these specific techniques was necessary to achieve such high performance. The system also proved to be efficient, capable of processing an image in a fraction of a second, which is fast enough to be useful in a busy hospital setting.

When compared to other advanced computer programs currently available, this new system outperformed them all. It beat the next best methods by a clear margin, showing that its unique approach to combining different types of visual information was the key to its success. The researchers noted that while the system is highly accurate, it was trained on data from a single hospital, meaning its ability to work on images from different machines or different patient populations still needs to be tested in the wider world. Nevertheless, the findings demonstrate that by giving artificial intelligence a way to look at medical images from multiple angles and with specialized attention to detail, it is possible to create tools that can assist doctors in making faster, more accurate assessments of pregnancy health. This work suggests that the future of prenatal care may involve computers that can see the invisible details of the placenta, helping to ensure safer outcomes for mothers and their babies.

Technical Summary: Dual-Encoder Cross-Scale Dependency Fusion Network for Placenta Segmentation

Problem Statement

Accurate segmentation of the placenta is critical for prenatal assessment, pregnancy management, and the diagnosis of placenta-related abnormalities. However, automatic segmentation remains challenging due to complex anatomical structures, ambiguous boundaries, intensity variations, and significant differences between imaging modalities. Traditional manual delineation is labor-intensive and subjective, while existing automated methods often struggle with:

Modality Limitations: Most approaches are designed for single imaging modalities (either MRI or Ultrasound), limiting adaptability across different imaging characteristics.
Structural Complexity: Placental images exhibit large morphological variations, unclear boundaries, and heterogeneous intensity distributions, which hinder conventional encoder-decoder networks from preserving fine structures.
Feature Interaction: Existing methods often lack effective mechanisms for modeling dependencies among multi-scale features, resulting in insufficient interaction between detailed spatial information and high-level semantic representations.
Boundary Preservation: Maintaining accurate boundary information during feature fusion remains difficult, particularly for medical images with ambiguous structures.

Methodology: DECSDF-Net

The authors propose a Dual-Encoder Cross-Scale Dependency Fusion Network (DECSDF-Net) designed to segment placenta regions in both Magnetic Resonance (MR) and Ultrasound (US) images. The architecture integrates a dual-encoder design with four specialized modules to enhance feature learning and boundary preservation.

1. Network Architecture

Dual-Encoder Design: The network employs two parallel encoding pathways to extract complementary hierarchical features. The input image is processed by a Cross-Directional Coordinate Attention Module (CDCAM) to enhance directional spatial information before entering the encoders. This allows the network to learn complementary representations from different feature pathways.
Progressive Decoder: Features from the two encoders are concatenated and refined through convolutional operations. High-level semantic features are upsampled and fused with higher-resolution encoder features to recover spatial details.

2. Key Modules

Cross-Directional Coordinate Attention Module (CDCAM): Designed to capture directional dependencies and strengthen boundary-related features. Unlike standard coordinate attention, CDCAM introduces explicit interaction between horizontal and vertical features. It aligns directional features, applies a shared 1×1 convolution, and uses learnable coefficients ( $\alpha, \beta$ ) to allow global contextual information from one branch to enhance the other, improving spatial localization.
Guided Cross-Scale Dependency Module (GCSDM): This module addresses the semantic gap between low-level spatial details and high-level contextual information. It aligns features from different scales (via down-sampling) and uses the deepest semantic feature ( $F_1$ ) as a guidance map to selectively enhance shallower features ( $F_2, F_3$ ). This is followed by channel attention and a residual connection to integrate multi-level information effectively.
Attention-Guided Enhanced Receptive Field Block (AGERFB): To handle irregular shapes and scale variations, AGERFB employs three parallel branches with different receptive fields (1×1, 3×3 with dilation 1, and 3×3 with dilation 2). It integrates channel attention (CAM) and spatial attention (SAM) to refine the fused features, enhancing discriminative responses for complex structures.
Progressive Multi-Scale Feature Fusion Module (PMSFFM): Located in the decoder, this module progressively aggregates features from different resolutions. It upsamples lower-resolution features and fuses them with higher-resolution features in a step-by-step manner, allowing semantic information to be gradually combined with spatial details to recover fine placental structures.

3. Training and Loss

The network is trained using Dice Loss to address the class imbalance between the foreground (placenta) and background. The model was optimized using the Adam optimizer with an initial learning rate of 0.001.

Datasets and Experimental Setup

The study utilized two in-house datasets constructed from Deyang People's Hospital:

Placenta-MR: 1,546 MR images (928 training, 309 validation, 309 testing).
Placenta-US: 1,557 Ultrasound images (935 training, 311 validation, 311 testing).
All images were manually delineated by experienced clinicians. Experiments were conducted on an NVIDIA RTX 4090 GPU.

Results

Quantitative Performance

DECSDF-Net outperformed seven state-of-the-art methods (including DualA-Net, CKDNet, A-Net, BMANet, LMFR-Net, PMFSNet, and DSNet) on both datasets.

Placenta-MR Dataset:
- Dice: 0.8838
- MCC: 0.8775
- Jaccard: 0.7923
- Improvement over the best competitor (A-Net): +0.0129 (Dice), +0.0135 (MCC), +0.0201 (Jaccard).
Placenta-US Dataset:
- Dice: 0.8778
- MCC: 0.8642
- Jaccard: 0.7826
- Improvement over the best competitor (DSNet): +0.0046 (Dice), +0.0049 (MCC), +0.0073 (Jaccard).

Ablation Studies

Ablation experiments on the Placenta-MR dataset confirmed the contribution of each module:

Baseline (Dual-encoder only): Dice 0.8496.
Baseline + PMSFFM: Improved to 0.8548.
Baseline + PMSFFM + CDCAM: Improved to 0.8562.
Baseline + PMSFFM + CDCAM + GCSDM: Improved to 0.8633.
Full Model (including AGERFB): Achieved the final Dice of 0.8838.
The results indicate that each module provides complementary improvements, with the full integration yielding the strongest performance.

Complexity and Efficiency

Parameters: 4.60 M.
GFLOPs: 17.96.
Inference Speed: 160.05 FPS.
The model achieves a favorable balance between high segmentation accuracy and computational efficiency, maintaining a lower parameter count than heavy competitors like CKDNet (59.34 M) and DSNet (29.33 M) while offering superior accuracy.

Optimizer Selection

To determine the optimal optimization strategy, five optimizers (Adam, Adamax, RMSprop, Rprop, and SGD) were evaluated on the Placenta-MR dataset. Adam achieved the best overall performance with Dice, MCC, and Jaccard scores of 0.8838, 0.8775, and 0.7923, respectively. RMSprop performed similarly but slightly lower, while Rprop and SGD yielded lower accuracy. Consequently, Adam was selected for all experiments.

Significance and Claims

The paper claims that DECSDF-Net provides accurate and robust placenta segmentation across both MR and Ultrasound modalities. The significance of the work lies in:

Dual-Modality Adaptability: Successfully addressing the challenge of segmenting placenta in both MR and US images, which have distinct visual characteristics.
Enhanced Feature Interaction: The introduction of GCSDM and CDCAM effectively models cross-scale dependencies and directional spatial relationships, overcoming limitations of standard encoder-decoder networks.
Clinical Potential: The method shows promising potential for automated prenatal assessment and computer-aided diagnosis, offering a tool that is more accurate and efficient than manual delineation.

The authors maintain a modest tone regarding future work, noting that current validation is limited to a single clinical center. They suggest that future research should focus on multi-center validation to assess robustness across diverse imaging devices and patient populations, as well as further architectural optimization for practical clinical deployment.