Multi-Task Boundary-Guided Decoder (MBGD)

• The paper demonstrates that MBGD enhances segmentation by integrating explicit boundary maps to refine both edge delineation and semantic mask prediction.
• MBGD features a dual-head architecture with a shared upsampling backbone, employing transposed convolutions and channel-wise concatenation for effective feature fusion.
• Empirical results show that integrating MBGD raises the Dice score by 1.39% and reduces the Hausdorff Distance by 3.39 mm in ultrasound image segmentation.

The Multi-Task Boundary-Guided Decoder (MBGD) is a neural network module tailored for medical image segmentation, designed to produce precise, spatially coherent masks by directly integrating boundary information into the semantic segmentation process. Developed in the context of the FreqDINO framework for ultrasound image segmentation, MBGD specifically addresses the challenge of boundary degradation in ultrasound images, leveraging a multi-tasking strategy to optimize both semantic mask accuracy and boundary delineation simultaneously (Zhang et al., 12 Dec 2025).

1. Architectural Role and Functional Overview

MBGD operates as the terminal decoding stage within FreqDINO, following the Frequency-Guided Boundary Refinement (FGBR) module. It receives a refined high-level feature tensor $$F_{\mathrm{refined}}\in\mathbb{R}^{B\times C\times H_1\times W_1}$$, where $B$ is the batch size, $C=512$ is the channel dimension, and the spatial resolution is typically $1/16$ of the input (e.g., $32\times32$ for $512\times512$ images). The fundamental purpose of MBGD is twofold:

• To generate a crisp, per-pixel boundary map $M_{\mathrm{boundary}}$ that accentuates anatomical edges
• To employ this boundary map for guiding the final semantic (mask) prediction $M_{\mathrm{mask}}$, ensuring alignment of predicted object borders with true anatomical structures

MBGD embodies a "boundary-first" decoding policy, in which explicit edge information is computed and expanded before influencing semantic masking.

2. Detailed Network Architecture

The MBGD architecture comprises a shared upsampling backbone and dual decoding heads:

1. Shared Upsampling Backbone: Four cascaded transposed convolutional "UpBlocks" ($\mathrm{kernel}=2$, $\mathrm{stride}=2$) progressively double the feature map resolution and reduce channels from $B$0 to $B$1. After four UpBlocks, the output $B$2 is at original resolution.
2. Dual Heads:
   • Boundary Head: Applies a $B$3 convolution to $B$4 to produce $B$5, followed by a sigmoid activation yielding per-pixel probabilities. This map is then lifted back to $B$6 feature dimensions via a $B$7 convolution: $B$8, resulting in $B$9.
   • Mask Head: Concatenates $C=512$0 and $C=512$1 along the channel axis (yielding $C=512$2), then collapses it via a $C=512$3 convolution to a scalar mask map $C=512$4. At evaluation, per-pixel mask probabilities are produced using a sigmoid.

This pipeline ensures that the mask head processes not just semantic content but explicit, refined boundary cues prior to making object-level predictions.

3. Mathematical Workflow and Multi-Task Objective

The MBGD pipeline leverages explicit mathematical operations:

• Upsampling: $C=512$5
• Boundary prediction: $C=512$6, $C=512$7
• Mask prediction: $C=512$8

Ground-truth binary boundaries $C=512$9 are computed from the mask ground truth $1/16$0 using the morphological gradient:

$1/16$1

Losses are computed as binary cross-entropy for both mask ($1/16$2) and boundary ($1/16$3), combined with a weighting coefficient $1/16$4:

$1/16$5

Both heads are optimized jointly from the outset under a fixed loss weighting, with the DINOv3 backbone weights frozen and only the adapters and modules in MFEA, FGBR, and MBGD updated. Adam optimizer is used (initial $1/16$6, decay factor $1/16$7 per epoch, batch size $1/16$8, training for $1/16$9 epochs).

4. Fusion Strategy for Enforcing Spatial Coherence

MBGD leverages a channel-wise concatenation mechanism for fusing boundary guidance into mask prediction. $32\times32$0 is derived by convolving the sigmoid-activated boundary map, encoding the spatial strength of edge confidences. This explicit concatenation, without additional attention mechanisms, enables the mask head to adaptively sharpen or soften mask predictions according to local boundary certainty. In empirical evaluations, this minimalist fusion sufficed to achieve improved contour accuracy.

5. Implementation Hyperparameters and Pseudocode

Component	Parameter	Value/Description
Input feature channels	$32\times32$1	512
Output feature channels	$32\times32$2	256
UpBlock	ConvTranspose2d	(in=256, out=256, kernel=2, stride=2)
Boundary head	Conv2d	(in=256, out=1, kernel=1)
Boundary feature conv	Conv2d	(in=1, out=256, kernel=3, padding=1)
Mask head	Conv2d	(in=512, out=1, kernel=1)
Loss weight	$32\times32$3	0.3
Optimizer	Adam	lr=$32\times32$4, decay=0.98/epoch
Training schedule	Epochs	300, batch=16
Framework		PyTorch, NVIDIA A5000, DINOv3-Large

Core pseudocode flow:

$32\times32$5

6. Empirical Performance and Ablation Analysis

Ablation studies establish the quantitative impact of MBGD within FreqDINO. When only the preceding MFEA and FGBR components are present, segmentation achieves Dice=85.13% and Hausdorff Distance (HD)=43.02 mm. Integrating MBGD raises Dice to 86.52% (+1.39%) and reduces HD to 39.63 mm (−3.39 mm). This demonstrates a substantial improvement in both segmentation overlap and boundary alignment, directly attributable to the boundary-guided mask decoding mechanism (Zhang et al., 12 Dec 2025).

7. Context and Methodological Implications

FreqDINO, containing the MBGD module, employs frozen foundation models (DINOv3) augmented with frequency-aware processing and explicit boundary refinement. MBGD’s design reflects a commitment to capturing fine-grained anatomical boundaries by multi-tasking mask and edge supervision. This approach aligns with broader trends in medical vision research, where dedicated boundary branches and auxiliary spatial losses are leveraged to counter the adverse effects of modality-specific imaging artifacts (e.g., speckle in ultrasound). The effectiveness of MBGD in the context of FreqDINO has implications for segmentation pipeline design in other domains where precise edge localization is critical.