DFF-Net: A Lightweight Dynamic Feature Fusion Network for Instance Segmentation of Axillary Brachial Plexus Ultrasound Images

Abstract

The brachial plexus nerves at the axillary level are small in size and have a tortuous course, intertwining with blood vessels and other tissues. Ultrasound images are affected by speckle noise and structural blurring, leading to errors in manual segmentation and reducing the accuracy and success rate of axillary nerve block localization. Current lightweight networks primarily focus on semantic segmentation, making it difficult to precisely distinguish the boundaries between adjacent nerve bundles and surrounding tissues. Additionally, they lack optimisation for axillary nerve instance segmentation, making it challenging to balance accuracy and real-time performance. To address these issues, we propose the Dynamic Feature Fusion Network, which enhances segmentation accuracy while maintaining lightweight architecture. This model combines a re-parameterised vision transformer with spatial pyramid pooling to compress parameter counts while maintaining feature expression capabilities. To achieve dynamic fusion of multi-scale features, this study proposes a SimAM-enhanced Bidirectional Fusion Network based on a similarity attention module, which improves segmentation accuracy in neurovascular intersection regions. Finally, a context-based mechanism is used to improve target localisation accuracy while completing pixel-level segmentation tasks, and a hybrid loss function is designed to enhance the model's localisation accuracy and optimise the training process. The method was evaluated on the Ultrasound Axillary Brachial Plexus dataset and experimental results showed that the proposed method achieved 57FPS and 0.604 mAP@0.5:0.95 on a single GPU. This significantly improves the segmentation accuracy and efficiency of the axillary brachial plexus, providing a reliable auxiliary tool for ultrasound-guided nerve block.

Introduction

The brachial plexus, formed by the ventral rami of cervical (C5–C8) and thoracic (T1) spinal nerves, gives rise to major peripheral nerves (median, radial, ulnar, musculocutaneous, etc.) that innervate the upper limb. At the axillary level, these nerves are densely packed within or adjacent to the neurovascular bundle, surrounded by muscles (e.g., coracobrachialis) and blood vessels (axillary artery/vein), presenting a complex anatomical environment [1]. Axillary brachial plexus block (ABPB) is a cornerstone of regional anesthesia for upper limb surgeries, where local anesthetics are injected around target nerves to block sensory-motor conduction. Compared to supraclavicular or infraclavicular approaches, ABPB avoids risks like pneumothorax or phrenic nerve palsy, making it preferred for outpatient and pediatric procedures [2]. However, its success hinges on accurate nerve localization: traditional blind techniques rely on anatomical landmarks or paresthesia, leading to a 10–20% failure rate due to individual anatomical variability [3]. Ultrasound-guided ABPB improves visibility, but challenges persist: (1) Imaging limitations: Speckle noise, low contrast between nerves and surrounding fascia, and overlapping structures (e.g., nerves vs. small veins) often obscure nerve boundaries. (2) Operator dependence: Manual segmentation and interpretation require years of expertise; even experienced anesthesiologists may misjudge nerve locations, risking intravascular injection (local anesthetic toxicity) or incomplete block (prolonged surgery, repeat punctures) [4]. (3) Time constraints: Intraoperative navigation demands real-time feedback to adjust needle trajectory, which is unmet by current labor-intensive manual workflows.

Currently, in the field of brachial plexus ultrasound image segmentation, segmentation techniques for regions such as the intermuscular groove [5] and supraclavicular fossa [6] have achieved certain results. Recent advances in AI-based medical ultrasound analysis have shown promise in addressing similar challenges. For instance, deep learning models have been developed to predict dynamic shoulder motion trajectories from ultrasound sequences, demonstrating AI’s potential in real-time procedural guidance [7]. In nerve segmentation, deep learning algorithms achieve robust performance in median nerve localization, with meta-analyzed Dice scores exceeding 0.85, underscoring the feasibility of AI-driven nerve segmentation in clinical practice [8]. However, axillary brachial plexus segmentation remains underexplored, with unique challenges: (1) Small, tortuous nerve bundles (2–5 mm in diameter) are easily confounded with adjacent vessels; (2) Multi-instance ambiguity (multiple nerves in a single field of view) requires distinguishing individual nerve instances; (3) Resource-constrained ultrasound devices demand lightweight models for on-edge deployment.

Existing instance segmentation methods, while successful in general medical imaging, fail to address these challenges: Heavy computational burden: Models like Mask R-CNN or SOLOv2 have millions of parameters and high FLOPs, resulting in inference speeds < 10 FPS, unsuitable for real-time navigation. Suboptimal feature fusion: Fixed-scale feature extraction struggles with the multi-scale nature of axillary nerves (e.g., larger median nerve vs. thinner musculocutaneous nerve), leading to blurred boundaries. Lack of anatomical adaptability: Generic loss functions (e.g., cross-entropy) do not account for the elongated, filamentous morphology of nerves, reducing segmentation precision in complex cases.

To bridge this gap, we propose DFF-Net, a lightweight dynamic feature fusion network specifically designed for instance segmentation of the axillary brachial plexus. This framework strikes a balance between real-time performance and segmentation accuracy, meeting the clinical demands of intraoperative guidance. We validate DFF-Net using the newly constructed Ultrasound of the Axillary Brachial Plexus (UABP) dataset, where it outperforms existing state-of-the-art methods. Our principal contributions are as follows:

• Clinically-Oriented Lightweight Design: By integrating a reparameterised RepViT model with fast spatial pyramid pooling, we reduce model parameters and floating-point operations compared to mainstream instance segmentation models, enabling real-time inference on portable ultrasound devices.

• Dynamic Multi-Scale Feature Fusion: An innovative Context-Feature Perception Module (CFPM) collaborates with a dynamically weighted feature pyramid to adaptively fuse multi-scale features. This significantly enhances the distinguishability of small nerves from adjacent vessels—crucial for mitigating intravascular injection risks.

• Anatomy-Aware Loss Optimisation: A composite loss function (dynamically weighted classification loss and boundary regression loss) targets the slender structural characteristics of nerves, improving instance boundary accuracy compared to standard loss functions.

Related Work

Ultrasound Image Segmentation of the Brachial Plexus

The integration of deep learning and ultrasound-guided nerve block techniques has significantly advanced the field of anaesthesia and pain management [9]. Yang et al. designed a U-Net variant to train a CNN to identify the brachial plexus at the intermuscular groove level [6]. Abraham et al. trained an M-Net model focused on brachial plexus recognition, optimising the U-Net structure to merge decoder outputs into the loss function and improve segmentation accuracy [10]. Wang et al. proposed an optimised and efficient ResU-Net variant for brachial plexus segmentation, combining dense convolutions and residual multi-scale pooling to reduce spatial information loss and enhance robustness across different scales [11]. Ding et al. designed a multi-object detection network, MallesNet, to enhance the accuracy and robustness of brachial plexus segmentation by leveraging auxiliary information from multiple objects [12]. Smistad et al. distinguished six categories of axillary brachial plexus nerves, with this fine-grained classification being more beneficial for identifying and locating target regions [13]. Inspired by the aforementioned work, the network structure proposed in this study improves identification accuracy at the axillary level through collaborative segmentation of multiple targets.

Medical Image Instance Segmentation

Medical image instance segmentation is a region-based segmentation method that performs pixel-level segmentation of human organs and lesion areas within each detection box. Huang et al. proposed that multi-instance learning can enhance the extraction and aggregation capabilities of different features [14]. Single-stage instance segmentation models simultaneously perform object detection and pixel-level segmentation in a single forward pass, emphasizing speed and efficiency, such as SSD [15] and YOLO [16]. An et al. designed an instance segmentation framework for fetal echocardiography heart chamber segmentation based on the SOLOv2 model, and developed a new category attention module to correct instance misclassification and improve segmentation accuracy [17]. Lin et al. designed improvements on YOLO, proposing a morphology-centric retinal ganglion cell instance segmentation algorithm for multi-dimensional glaucoma analysis [18]. Liu et al. designed a contour-based instance segmentation method for segmenting separable villi in placental tissue pathology images and developed a point-wise size-adaptive scoring network to measure the quality of contour localization [19]. Two-stage instance segmentation models, which separate detection and segmentation tasks and combine multi-scale feature optimisation, can better handle complex scenes and fine structures in medical images, such as the commonly used Mask RCNN [20] and its variants [21–23], but they have slower inference speeds. This study addresses medical segmentation tasks with limited hardware resources by comprehensively considering strategies that balance accuracy and speed, selecting a single-stage model as the base method, and combining feature fusion strategies for optimisation and accuracy improvement.

Methods

The overall architecture of our proposed DFF-Net is illustrated in Fig. 1, which clearly delineates its two core components: the Feature Extraction Network and the Dynamic Feature Fusion Module.

Fig. 1.

DFF-Net network structure used for axillary brachial plexus segmentation

• Feature Extraction Network: The backbone of DFF-Net is a lightweight Vision Transformer, RepViT [24], which efficiently extracts multi-scale feature maps from an input image. Following the RepViT stages (Stem, Stage 1–4), we incorporate a Spatial Pyramid Pooling Fast (SPPF) module to further enrich the feature representation at different scales. This process yields a hierarchy of feature maps, denoted as C2, C3, C4, and C5, where lower levels (e.g., C2) retain finer spatial details and higher levels (e.g., C5) carry stronger semantic information.

• Dynamic Feature Fusion Module: To address the challenge of segmenting small and intertwined nerves, we introduce a dedicated fusion module. The multi-scale features (C3, C4, C5) from the backbone are fed into a Similarity-based Bidirectional Feature pyramid network (SimBiF). Within SimBiF, the Contextual Feature Perception Module (CFPM) enhances the features' discriminative power, while the Dynamic Sample module (DySample) performs adaptive upsampling to recover spatial details. A key design is the implicit fusion of the high-resolution C2 features into the P3 level via upsampling, effectively balancing detail and semantics. This module outputs a set of optimized, fused feature maps P3, P4, and P5.

• Segmentation Head: The refined features (P3-P5) are then directed to the segmentation head. Here, we employ a prototype-based mask generation strategy that collaboratively uses these multi-level features to simultaneously predict bounding boxes, class probabilities, and high-fidelity, pixel-level segmentation masks.

Re-parameterised Multi-scale Feature Extraction Network

The feature extraction network adopts the RepViT, a lightweight visual transformer with reparameterisation. This module improves training efficiency through a lightweight architecture design combined with a multi-branch topological structure. To better adapt to the needs of low-computing-power devices and dense instance segmentation tasks, we modified and optimised the feature extraction network based on the original structure. The specific structure is shown in Fig. 2. In the Stem layer, two 3 × 3 convolutions with a stride of 2 are used to achieve efficient downsampling and feature preprocessing. The main part consists of four stages, each of which is composed of multiple stacked RepViTBlocks and RepViTSEBlocks. Shallow feature extraction primarily uses RepViTBlocks to preserve local details, while deep feature extraction primarily uses RepViTSEBlocks to enhance the channel selection capability of semantic features.

Fig. 2.

Feature extraction network RepViT structure

In the RepViTSEBlock module, the network retains the RepViT's reparameterised structure while optimising its design, ensuring that it remains lightweight while improving performance. During training, a multi-branch structure is employed: the main branch uses 3 × 3 convolutions to capture spatial details (such as edges and textures), while the auxiliary branch's 1 × 1 convolutions enhance feature fusion between channels, complementing patterns not covered by the main branch. Residual connections preserve the original input to stabilise gradients. During inference, based on the additivity and equivalent mathematical transformations of convolutional and BN layers, they are merged into a single 3 × 3 DW convolution. During merging, 1 × 1 convolutions are expanded to 3 × 3 kernels via zero-padding, with the centre value being the original weight and surrounding values set to 0. Residual connections are expanded to 3 × 3 convolution kernels via padding with an identity matrix (diagonal elements are 1, others are 0).

Additionally, a reparameterised deep separable convolution (RepVGGDW) is designed as a token mixer for segmentation tasks, combined with a channel mixer using an inverted residual structure to balance local feature extraction and cross-channel interaction. Furthermore, by combining convolution and Squeeze-Excite (SE) modules, the model dynamically adjusts channel features and adapts to dense task requirements through configurable channel counts and downsampling positions. These modifications significantly reduce the computational complexity of model training, improve inference speed, and perform particularly well on low-computing-power devices.

Multi-scale Feature Dynamic Fusion Network

Dynamic Adaptive Feature Pyramid Network SimBiF

To address the characteristics of medical ultrasound images, such as low contrast, speckle noise, and blurred boundaries, we designed a dynamic adaptive feature pyramid network, SimBiF, for cross-scale feature fusion. The SimBiF network employs a dynamic feature weighting mechanism combined with depth-separable convolutions to reduce computational complexity. It is specifically optimised for the multi-scale feature fusion requirements of medical ultrasound images, retaining the original bidirectional path structure of BiFPN [25] to support multi-scale feature fusion. By integrating high-level semantic information and low-level details through bidirectional information flow, it achieves multi-scale feature synergy, overcoming the limitations of traditional BiFPN’s fixed-weight fusion. The specific structure is shown in Fig. 3.

Fig. 3.

SimBiF structure of dynamic adaptive feature pyramid network

To achieve fine-grained enhancement of ultrasound image features while adaptively suppressing speckle noise, we incorporate the SimAM [26] attention mechanism into feature map preprocessing. SimAM generates pixel-level attention maps by computing an energy function based on neural similarity, thereby amplifying feature focus in low-contrast regions and enabling dynamic pixel-level weighting. The calculation method for the attention weights of the i-th pixel is described in Eq. 1, where k is the normalisation constant, $N_i$ is the set of neighbouring pixels of the i-th pixel, and $s\left(f_i,,,f_j\right)$ is the similarity between the i-th pixel and the j-th pixel.

$$\begin{array}{c}{W}_i=\frac{1}{k}\sum _{j\in N_i}s\left(f_i,,,f_j\right)\end{array}$$ 1

The attention map $W_i$ generated by SimAM enhances the feature map P prior to its entry into bidirectional pyramid fusion, as shown in Eq. 2. Here, the $\odot$ denotes the Hadamard product, i.e., element-wise multiplication. This process dynamically increases the weighting of significant, information-rich targets or boundary regions within the image while adaptively suppressing the weighting of speckle noise and low-information background areas, thereby accentuating target features.

$$\begin{array}{c}\overline{P}=P\odot W_i\end{array}$$ 2

The specific process of feature fusion at different scales is described as follows. For feature map P5, in the top-down path, it is upsampled to obtain ${P5}_{\mathit{td}}$, which is then weightedly fused with feature P4. After 1 × 1 convolution to adjust the channels, ${P4}_{\mathit{td}}$ is obtained. The implementation method is described in Eq. 3, where ${\omega }_4$ and ${\omega }_5$ are the feature attention weights, respectively:

$$\begin{array}{c}{P4}_{\mathit{td}}=Conv1x1\left(\frac{{\omega }_4*\overline{P4}+{\omega }_5*Upsample\left(\overline{P5}\right)}{{\omega }_4+{\omega }_5+\epsilon }\right)\end{array}$$ 3

In the bottom-up path, the feature map ${P4}_{\mathit{out}}$ is downsampled and then weightedly fused with P5. Finally, a 1 × 1 convolution is performed to obtain the final feature map ${P5}_{\mathit{out}}$. The specific process is described in Eq. 4:

$$\begin{array}{c}{P5}_{\mathit{out}}=Conv1x1(\frac{{\omega }_4^{\prime }*Downsample(\overline{{P4}_{\mathit{out}}})+{\omega }_5*\overline{P5}}{{\omega }_4^{\prime }+{\omega }_5+\epsilon })\end{array}$$ 4

The same mechanism is used for feature map P4 to obtain the final output ${P4}_{\mathit{out}}$, as shown in Eq. 5. Specifically, ${{\omega }_4}^{\prime \prime }$ denotes the independent, learnable scalar fusion weight applied to the original input feature $\overline{\text{P}4}$ after weighted enhancement by SimAM, to distinguish it from the weight ${\omega }_4\prime$ derived from the $\overline{{\text{P}4}_{\mathit{td}}}$ pathway:

$$\begin{array}{c}{P4}_{\mathit{out}}=Conv1x1(\frac{{\omega }_3^{\prime }*Downsample(\overline{{P3}_{\mathit{out}}})+\omega _4^{_{}^{\prime \prime }}*\overline{P4}+{\omega }_4^{\prime }*\overline{{P4}_{\mathit{td}}}}{{\omega }_3^{\prime }+\omega _4^{_{}^{\prime \prime }}+{\omega }_4^{\prime }+\epsilon })\end{array}$$ 5

To meet the requirements for multi-scale features in segmentation tasks, bidirectional fusion is used to provide comprehensive scale information. The output features of each layer are then sequentially output upwards, i.e. from high to low resolution, to subsequent modules. To address the issue of scale changes in processing tasks, this study employs bilinear interpolation to introduce dynamic scaling for aligning feature maps of different resolutions. Additionally, cross-scale connections with adaptive weights are added to ensure robust feature fusion across a wide range of scales.

Lightweight Dynamic Upsampling Operator DySample

Feature upscaling is a critical component in segmentation task models. The most commonly used samplers, such as nearest neighbour and bilinear interpolation, follow fixed rules and lack flexibility. Deformable convolutions achieve adaptive sampling through global offset learning, but this increases computational complexity and may lead to detail loss. To handle more complex feature distributions and enhance detail retention in segmentation tasks, we propose the DySample module inspired by point sampling [27], which uses dynamically learned offsets to adaptively adjust the precise offset for each pixel point.

The DySample module predicts pixel-level offsets through lightweight convolution layers, then generates $s$ equidistant coordinate points as sampling points by predicting the relative offsets of the input grid. Through Cartesian product operations, the initialised position $P_{\mathit{init}}$ is obtained, as defined in Eq. 6, where $s$ is the scaling factor.

$$\begin{array}{c}{P}_{\mathit{init}}=\left\{\left(\frac{2i-s+1}{2s},,,\frac{2j-s+1}{2s}\right),|i,j\in ,\left\{0,,,1,,,\cdots ,,,s,-,1\right\}\right\}\end{array}$$ 6

To accommodate different task types, two distinct forms of upsampling are provided during offset generation. For real-time tasks, offsets $\mathrm{\Delta }p$ can be directly predicted at low resolution. The process involves predicting offsets via a 1 × 1 convolution in the original resolution space, with a scaling factor of 0.25 constraining offsets within the range [− 0.5, 0.5] pixels to prevent sampling points from spanning adjacent feature units. For dense tasks, the input features are first upsampled using PixelShuffle [28], then the offset $\mathrm{\Delta }p$ is predicted in the high-resolution space, and finally the sampling grid is adjusted using PixelUnshuffle. The implementation is shown in Eq. 7, where $f_{\mathit{offset}}$ is the 1 × 1 convolution layer and $X$ is the input feature.

$$\begin{array}{c}\mathrm{\Delta }p=\left\{\begin{array}{c}{f}_{\mathit{offset}}\left(X\right).0.25+P_{\mathit{init}}\left(L,P\right)\\ PixelUnshuffle\left(f_{\mathit{offset}},\left(P,i,x,e,l,S,h,u,f,f,l,e,\left(X,,,s\right)\right),,,s\right)\bullet 0.25+P_{\mathit{init}}\left(P,L\right)\end{array}\right)\end{array}$$ 7

The specific sampling process is shown in Eq. 8, $\widehat{X}\left(p\right)$ represents the upsampled feature value obtained at the target coordinate p on the output feature map following dynamic sampling and bilinear interpolation operations. It corresponds to the value of our DySample module's output feature map c at that position. $X(q)$ represents the value of the input feature at coordinate $q$, $p$ is the target coordinate of the output feature map, $\omega$ is the bilinear weight, and $N(p+\mathrm{\Delta }p)$ is the four nearest neighbour pixel coordinates after the output coordinate p is mapped back to the input feature map.

$$\begin{array}{c}\widehat{X}\left(p\right)={\sum }_{q\in \mathbb{N}\left(p,+,\mathrm{\Delta },p\right)}X\left(q\right)\bullet \omega \left(p,+,\mathrm{\Delta },p,,,q\right)\end{array}$$ 8

Context Feature Perception Module CFPM

To compensate for the insufficient feature representation capabilities of feature extraction networks caused by lightweight design and to improve the completeness and boundary accuracy of objects in segmentation tasks, we designed the Context-aware Feature Perception Module (CFPM). This module enhances the ability to capture multi-scale contextual information through a dual-branch design combining the up-down guidance module [29] and dilated convolutional context, followed by dynamic feature fusion via global attention. Subsequently, the efficient mobile bottleneck convolution MBConv is employed to improve computational efficiency, making the model more suitable for real-time processing scenarios.

The specific structure of CFPM is shown in Fig. 4. The input features are first expanded by a 1 × 1 convolution in MBConv to increase the number of channels in the feature map, thereby providing more feature representation space for subsequent operations. We use depth-separable convolutions to reduce computational complexity while preserving the spatial characteristics of the information. Finally, a 1 × 1 compression convolution is applied to reduce the number of channels back to a smaller value for the final output. In the Bottleneck module, local features and global contextual information are fused, normalised, and activated before passing through a global feature guidance module. This combination of local features and surrounding contextual information enhances the expressive power of the features. Additionally, we use channel-level convolutions to extract local information from the input feature map, focusing on learning local spatial structures within each channel. Finally, we process this global information through two fully connected layers and fuse it with the original features to enhance the global consistency of the features.

Fig. 4.

Structure of the context-aware module CFPM

Composite Loss Function Design

For collaborative training of multiple tasks in instance segmentation, we adopted a composite loss. As shown in Eqs. 9 to 10, the overall loss consists of three parts: classification, localisation, and segmentation, and gradient equilibrium is achieved through adaptive weighting.

$$\begin{array}{c}{L}_{\mathit{total}}={\omega }_{\mathit{cls}}{*L}_{\mathit{cls}}+{\omega }_{\mathit{bbox}}{*L}_{\mathit{bbox}}+{\omega }_{\mathit{seg}}{*L}_{\mathit{seg}}\end{array}$$ 9

$$\begin{array}{c}{\omega }_{\mathit{bbox}}*\mathrm{\nabla }{L}_{\mathit{bbox}}\approx {\omega }_{\mathit{cls}}*\mathrm{\nabla }{L}_{\mathit{cls}}\approx {\omega }_{\mathit{seg}}*\mathrm{\nabla }{L}_{\mathit{seg}}\end{array}$$ 10

For the coupled classification-localisation task, we employed an innovative dynamic weighted classification loss function, SlideLoss-IoU [30], which retains the original binary cross-entropy loss calculation method but enforces element-wise computation. This mechanism dynamically adjusts the loss weights based on the IoU values and target scores between the predicted and ground truth bounding boxes, effectively balancing the relationship between the classification and localisation tasks. As shown in Eq. 11, SlideLoss-IoU divides samples into three regions based on their IoU values. For low-quality regions, a fixed weight is used to maintain the baseline gradient. For critical regions, exponential weights and adaptive threshold reinforcement learning are employed. For high-quality regions, exponential decaying weights are used for fine-tuning.

$$\begin{array}{c}f\left(x\right)=\left\{\begin{array}{c}1x\le \mu -0.1\\ e^{1-\mu }\mu <x<\mu -0.1\\ e^{1-x}x\ge \mu \end{array}\right)\end{array}$$ 11

Segmentation loss introduces key improvements to the traditional BCE loss by incorporating bounding box-guided mask clipping and loss normalisation, making it more suitable for instance segmentation tasks. Boundary regression loss combines WIoU loss [31] for localisation loss and distribution focus loss DFL [32].

Compared to traditional IoU variants, the WIoU loss automatically adjusts the weights of samples of different qualities by introducing a dynamic focus factor. It dynamically adjusts the gradient gain to assign more reasonable gradient weights to low-quality samples, such as small targets or severely occluded targets, thereby avoiding training being dominated by simple samples. The corresponding formulas are explained in Eqs. 12 to 13, where the base IoU loss $L_{\mathit{IoU}}=1-IoU$, ${\gamma }_{\mathit{dyn}}$ is the dynamic focus loss, and $\delta =IoU-{\mathit{IoU}}_{\mathit{mean}}$. Introducing DFL into the bounding box regression loss enables the capture of subtle changes in coordinates through probability distribution modelling, significantly improving bounding box localisation accuracy, particularly for small targets and blurry edges. The corresponding formula is explained in Eq. 14, where $S$ is the discrete probability distribution predicted by the model, and y is the target distribution.

$$\begin{array}{c}{L}_{\mathit{WIoU}}=\frac{{\gamma }_{\mathit{dyn}}*L_{\mathit{IoU}}}{mean\left(L_{\mathit{IoU}}^{*}\right)}\end{array}$$ 12

$$\begin{array}{c}{\gamma }_{\mathit{dyn}}=\frac{\delta }{\delta +\mathrm{\Delta }}\end{array}$$ 13

$$\begin{array}{c}DFL\left(S,,,y\right)=-{\sum }_{i=0}^{reg\_max-1}{y}_i\mathit{log}\left(S_i\right)\end{array}$$ 14

Experiments

Dataset

This study constructed the UABP dataset, focusing on the segmentation of the axillary brachial plexus, which was collected by the Affiliated Cancer Hospital of Chongqing University.

Data Collection

During data collection, high-frequency probes from three different ultrasound devices—Canon, TOSHIBA, and Mindray—were placed perpendicular to the long axis of the upper arm at the junction of the axillary fold and the biceps brachii muscle; during scanning, the axillary artery and vein, which are located medial to the biceps brachii and coracobrachialis muscles, were relatively easy to identify, and the median nerve, ulnar nerve, and radial nerve surrounding the axillary artery were localized by moving the probe proximally along the limb and tilting it appropriately, with the musculocutaneous nerve situated between the biceps brachii and coracobrachialis muscles. A total of 217 videos (average duration: 8–10 s per video) were collected in this study, distributed across devices as follows: Canon (n = 75), TOSHIBA (n = 68), and Mindray (n = 74); these videos originated from 185 distinct subjects (e.g., 100 healthy subjects and 85 patients), with some subjects (such as 32) providing multiple video segments due to multi-site examinations or repeated measurements.

Subjects who voluntarily participated and provided signed informed consent were included if their ultrasound image quality was deemed adequate for manual annotation and analysis. Those with local anatomical abnormalities, severe skin conditions affecting the imaging area, pregnancy, or inability to cooperate with the ultrasound examination were excluded. The recruited cohort comprised a diverse group of 185 subjects (including patients and healthy volunteers) with an average age of 67 ± 13 years (age range: [13, 67] years).

Data Preprocessing

After data collection, this study employed a key frame extraction method based on inter-frame difference analysis and local extreme value detection to screen the dataset. A total of 3–4 frames could be extracted from each ultrasound video, which were screened to ensure that at least one target was present in the selected frames, with blurred images removed. Each ultrasound image includes anatomical structures and corresponding annotations: three specialist physicians from the Department of Anaesthesiology at Chongqing University Cancer Hospital manually labeled the images in strict accordance with anatomical definitions from Peripheral Nerve Blockade and Anatomy for Ultrasound-Guided Regional Anaesthesia (3rd edition) [33]. All annotations were subsequently independently reviewed by a senior consultant; where discrepancies arose regarding boundary or structural identification, a third consultant was introduced for arbitration and final confirmation to ensure gold standard accuracy. Some images and visualisation masks in the UABP dataset are shown in Fig. 5. The key anatomical structures are denoted as follows: axillary artery (AA), musculocutaneous nerve (MCN), median nerve (MN), ulnar nerve (UN), and radial nerve (RN). Due to factors such as device model, probe angle, and patient variability, the original ultrasound data may exhibit insufficient sample diversity and high intraclass variability. This study employed letterbox image preprocessing techniques for uniform scaling and normalisation based on mean and standard deviation. Additionally, horizontal flipping was used to simulate symmetric probe scanning from both sides of the human body, and speckle noise was added to mimic imaging noise characteristics of different ultrasound devices, thereby enhancing the model's robustness to device variations.

Fig. 5.

Examples of arteries and nerves in ultrasound images. a shows raw ultrasound images collected from different patients. b shows basic markings made under the guidance of a specialist physician

Data Set Division

The original data in this data set is divided into a training set and a validation set in a 7:3 ratio. The augmented data is added to the training set to prevent overfitting. The final data set includes 2,157 ultrasound images and corresponding annotation files, of which the training set includes 1,941 ultrasound images and corresponding annotation files.

Implementation Details

In this section, the segmentation results of nerves and their anatomical structures are discussed through various experiments. All results were evaluated on the UABP dataset. The proposed network was implemented using the PyTorch 1.10.0 framework and ran on an NVIDIA RTX 3080 Ti GPU. All images were resized to 512 × 512, with 16 images per GPU. The model was optimised using the Adam optimiser, with a learning rate of 0.001, momentum of 0.9, and a maximum training epoch limit of 150.

Evaluation Metric

For the task of segmenting the brachial plexus in the axilla, the model needs to identify and segment multiple target instances in the image, each of which may belong to a different category. The confusion matrix is used to compare the relationship between the model's predicted instances and the true labels. Based on this evaluation framework, the area under the Precision-Recall curve (AP) is calculated for each category, while mAP is the average of all category APs, as described in Eq. 15. Using mAP@0.5:0.95 as the primary metric, this metric calculates the average precision mean at IoU thresholds ranging from 0.5 to 0.95 (with a step size of 0.05), serving as a comprehensive performance metric for the model across different localization resolutions. The definition formula is outlined in Eq. 16.

$$\begin{array}{c}AP={\int }_0^{1}p\left(r\right)dr\end{array}$$ 15

$$\begin{array}{c}mAP@0.5:0.95=\frac{1}{10}{\sum }_{IoU=0.5}^{0.95}{\mathit{AP}}_{\mathit{IoU}}\end{array}$$ 16

This study also used the APₛ (small target), APₘ (medium target), and APₗ (large target) analysis models to analyse the adaptability of brachial plexus cross-sections at different scales.

Recall is defined as the ratio of correctly detected positive instances to all actual positive instances, measuring the model’s ability to identify all relevant nerve structures:

$$\begin{array}{c}Recall=\frac{TruePositives}{TruePositives+FalsePositives}\end{array}$$ 17

FPS (Frames Per Second) measures the model’s inference speed, defined as the number of images processed per second, critical for real-time intraoperative navigation.

Statistical Analysis

For the two experiments described in the sections on “ablation studies” and “comparative studies,” statistical analysis of the results employed mAP@0.5:0.95 across all categories as the primary metric for evaluating segmentation accuracy. We conducted detailed statistical significance analyses for all variables requiring comparison in Tables 1, 2, and 3. Replications were performed using N = 10 distinct random seeds, with results reported as mean ± standard deviation. Performance differences between models were assessed using the paired Wilcoxon signed-rank test at a significance level of α = 0.05. The symbol † indicates statistically significant performance improvement at the p < 0.05 level relative to the configuration immediately above it (i.e. prior to the newly added component). The symbol ** indicates statistically significant improvement at the p < 0.01 level for the final DFF-Net configuration relative to the best-performing configuration among all others.

Table 1.

Ablation experiment results of the DFF-Net model on the UABP dataset

Model	mAP@0.5:0.95						Param (M)	FLOP (G)
	AA	MCN	MN	RN	UN	ALL
Baseline	0.723	0.505	0.504	0.397	0.401	0.506 ± 0.002	3.26	12.0
+ RepViT	0.732	0.487	0.497	0.428	0.448	0.518† ± 0.001	4.38	15.7
+ RepViT + SimBiF	0.728	0.484	0.502	0.437	0.429	0.528† ± 0.002	4.19	16.0
+ RepViT + DySample	0.738	0.493	0.511	0.451	0.449	0.527† ± 0.001	4.39	15.8
+ RepViT + CFPM	0.721	0.499	0.483	0.395	0.365	0.493 ± 0.002	3.94	14.8
+ RepViT + SimBiF + DySample	0.734	0.509	0.517	0.468	0.431	0.530† ± 0.001	4.20	16.0
+ RepViT + DySample + CFPM	0.713	0.476	0.513	0.442	0.416	0.512 ± 0.002	3.93	14.8
+ RepViT + SimBiF + CFPM	0.724	0.503	0.506	0.402	0.407	0.507 ± 0.002	3.79	15.0
Ours	0.745	0.513	0.521	0.484	0.459	0.545** ± 0.001	3.79	15.0

Table 2.

Ablation comparison results of the DFF-Net model on the UABP dataset with different loss functions

Loss function	mAP@0.5:0.95						Recall
	AA	MCN	MN	RN	UN	ALL
CIoU	0.745	0.513	0.521	0.484	0.459	0.545 ± 0.001	0.895 ± 0.004
SlideLoss + CIoU	0.749	0.519	0.575	0.496	0.531	0.576† ± 0.002	0.917† ± 0.003
SlideLoss + ECIoU	0.737	0.508	0.527	0.483	0.471	0.546 ± 0.003	0.924 ± 0.002
SlideLoss + PIoU	0.74	0.543	0.532	0.501	0.487	0.561 ± 0.002	0.925 ± 0.004
SlideLoss + SIoU	0.752	0.509	0.537	0.489	0.5	0.558 ± 0.003	0.937 ± 0.002
SlideLoss + MDPIoU	0.685	0.403	0.421	0.331	0.234	0.413 ± 0.005	0.697 ± 0.003
Ours	0.764	0.563	0.585	0.554	0.552	0.604** ± 0.001	0.952** ± 0.001

Table 3.

Comparison of different instance segmentation models on the UABP dataset

Model	mAP@0.5	mAP@0.5:0.95	APL	APM	APS	Param(M)	FPS
Mask-RCNN	0.968	0.578 ± 0.004	0.744	0.581	0.448 ± 0.003	63.49	16
CondInst	0.618	0.289	0.678	0.293	0.247	58.3	18
SOLOv2	0.604	0.274	0.625	0.28	0.127	65.69	12
FastInst	0.509	0.214 ± 0.003	0.606	0.219	0.099 ± 0.002	34.8	22
BoxInst	0.661	0.166	0.154	0.17	0.135	53.24	14
BlendMask	0.901	0.478	0.746	0.479	0.434	55.18	18
Yolov8	0.974	0.515 ± 0.002	0.824	0.514	0.398 ± 0.002	3.26	64
Ours	0.982	0.604** ± 0.001	0.853	0.603	0.477** ± 0.002	3.79	57

Results

Ablation Studies

To evaluate the effectiveness of the proposed DFF-Net, ablation studies were conducted by adding or removing one or more modules from the proposed method. To reveal the different performances of each module, all settings were kept the same except for the specific changes defined in each experiment (following the training procedure described in “Implementation Details”). The results of the different ablation experiments are shown in Table 1.

Ablation results demonstrate that the modules of DFF-Net contribute distinct and significant performance gains. The integration of RepViT improves overall mAP by 1.2% (†, p < 0.05) with a moderate increase in computational cost. The addition of SimBiF further enhances accuracy by 1.0% (†, p < 0.05), while DySample (Dynamic Sampling Module) also contributes positively to performance. In contrast, CFPM (Contextual Feature Perception Module) reduces the number of parameters by 10.1% while maintaining strong performance on large objects. The combination of RepViT, SimBiF, and DySample yields the best performance among partial configurations, achieving an mAP of 0.530. Notably, the full model (Ours), which integrates RepViT, SimBiF, DySample, and CFPM, achieves the highest overall mAP of 0.545—a 7.7% improvement (**, p < 0.01) over the best partial configuration—while maintaining parameter efficiency (3.79 M) and computational cost (15.0G FLOPs). This synergistic improvement validates the design of each module and their collective role in addressing distinct challenges in feature extraction, fusion, and efficiency for axillary brachial plexus segmentation.

To further visually observe the segmentation performance of DFF-Net across different categories, we utilise a normalised confusion matrix for presentation. As shown in Fig. 6, this study includes five categories, so the confusion matrix will be a 6 × 6 matrix (including the background class), where rows represent the true categories and columns represent the predicted categories. The values in the matrix represent the normalised classification probabilities. Observing the values along the diagonal reveals that DFF-Net achieves high classification accuracy across most categories. The off-diagonal elements of the confusion matrix identify the confusion between different categories. Although DFF-Net exhibits some confusion between certain categories, its overall classification performance across different categories is satisfactory.

Fig. 6.

Confusion matrix of DFF-Net on UABP dataset

Additionally, this study evaluated the effectiveness of each component of the composite loss function. We conducted ablation studies on the model using the UABP dataset. We systematically assessed the performance of different combinations of loss functions in object detection tasks, with a focus on the effectiveness of SlideLoss in conjunction with various IoU variants such as WIoU, ECIoU [34], PIoU [35], SIoU [36], and MDPIoU [37].

The experimental results of loss function ablation are presented in Table 2. As shown, the combination of SlideLoss and WIoU (Ours) significantly outperforms all other loss configurations, achieving the highest overall mAP of 0.604 and recall of 0.952—both statistically significant at the p < 0.01 level (**). Compared to the baseline CIoU, the proposed loss combination improves overall mAP by 10.8% (from 0.545 to 0.604), with particularly notable gains on challenging categories such as RN and UN, which improve by 14.5% and 20.3%, respectively. In addition, the proposed method also surpasses other SlideLoss-based variants (e.g., + ECIoU, + PloU, + SloU) by a clear margin in overall mAP, demonstrating the unique compatibility between SlideLoss and WIoU. These results validate that SlideLoss enhances feature discrimination while WIoU provides adaptive localization weighting, and their joint use exhibits a clear synergistic effect, effectively improving both classification reliability and bounding box accuracy in axillary brachial plexus segmentation.

To enhance the clarity of the results and facilitate analysis, we utilise a polygon radar chart to present the results of the loss function ablation comparison experiment. As shown in Fig. 7, the vertices of the radar chart represent different categories of metrics, with metrics closer to the vertices indicating higher values. This provides an intuitive comparison of the performance distribution of different loss function combinations across six evaluation metrics. Among these, the coverage area of SlideLoss + WIoU is significantly larger than that of other methods, particularly in low-baseline categories such as RN and UN, where its contour nearly encompasses all comparison methods, validating the comprehensive superiority of this combination.

Fig. 7.

Polygonal radar chart of different loss function metrics

Comparative Studies

In this study, we evaluated the performance of DFF-Net against several state-of-the-art instance segmentation methods, including the classic two-stage Mask R-CNN, as well as more recent models such as CondInst [38], SOLOv2 [39], FastInst [40], BoxInst [41], BlendMask [42], and YOLOv8 [43]. As shown in Table 3, our method consistently outperforms all competing approaches across multiple evaluation metrics, particularly in mAP@0.5, mAP@0.5:0.95, and segmentation performance across object scales, demonstrating high accuracy and robustness.

To rigorously assess the significance of these improvements, we conducted statistical comparisons between DFF-Net and three strategically selected representative models: YOLOv8, as the leading lightweight model for real-time segmentation; Mask R-CNN, representing classical architectures; and FastInst, a recently proposed efficient model for medical imaging. These models were chosen to validate specific advantages of DFF-Net—real-time capability, dynamic feature fusion superiority, and anatomical awareness—while avoiding multiple testing bias from exhaustive pairwise comparisons. As shown in Table 3, when testing using fundamental mAP@0.5:0.95 metrics and APs suitable for small-sample tests, DFF-Net achieves a statistically significant improvement over all three models at the p < 0.01 level (**), confirming the effectiveness of our proposed contributions.

Notably, DFF-Net excels in segmenting axillary brachial plexus structures in ultrasound images, achieving the highest mAP score while maintaining competitive inference speed, making it a suitable choice for clinical applications requiring both accuracy and efficiency.

To visually observe the overall accuracy of different models, we used a multi-indicator bar chart to show the performance comparison of multiple instance segmentation models under normalised indicators, with the maximum value of the indicator across all models highlighted. As shown in Fig. 8, each coloured bar represents a different metric, with taller bars indicating higher metric values. It can be observed that DFF-Net has the tallest bars across all metrics, particularly in mAP@0.5 and mAP@0.5:0.95, indicating its high accuracy and robustness in detection and segmentation tasks.

Fig. 8.

Comparison of histograms of segmentation metrics for different instance segmentation models

To better illustrate the balance between accuracy and speed across different models, we use a speed-accuracy scatter plot for visualisation. As shown in Fig. 9, each scatter point of a different colour represents a model, with the size and shade of the scatter point indicating the model's parameter count. The x-axis denotes the model's FPS on a single GPU, while the y-axis represents the model's overall accuracy on the UABP dataset. As can be seen from the figure, Mask-RCNN and DFF-Net perform best on the mAP@0.5:0.95, but Mask-RCNN has slower inference speed, while DFF-Net maintains high performance with faster inference speed and fewer parameters.

Fig. 9.

Scatter plot comparing accuracy and speed of different models on the UABP dataset

The visualisation segmentation results on the test set are shown in Fig. 10. Specifically, Fig. 10a shows the Ground Truth after label visualisation. Noise in the image produces a grainy texture, blurring the boundaries of anatomical structures, hindering visual perception, and reducing the quality of ultrasound images. Figure 10b–i present the visualisation results of each model. It can be observed that the segmentation results of DFF-Net, Mask-RCNN, and Yolov8n-seg are generally close to the true results. While BoxInst can determine the target location, the contours are not sufficiently clear. SOLOv2 and FastInst demonstrate advantages in segmenting large targets but are less accurate in identifying small targets.

Fig. 10.

Training results using different state-of-the-art methods on the UABP dataset. a Label visualisation, b–h The segmentation results of other advanced models. i The segmentation results of our network

Discussion

This study demonstrates that the proposed DFF-Net plays a crucial role in the segmentation of anatomical structures (nerves, arteries) in the axillary brachial plexus region. The DFF-Net, which combines RepViT with spatial pyramid pooling to form a lightweight multi-scale feature extraction architecture, effectively captures the features of neural targets of different sizes, significantly improving segmentation accuracy while drastically reducing the number of model parameters, thereby meeting the requirements of resource-constrained devices.

Within the current research landscape, most instance segmentation models (such as Mask RCNN and BlendMask) typically rely on heavyweight backbone networks. These networks sacrifice inference speed for accuracy, rendering them unsuitable for real-time ultrasound-guided procedures. Conversely, existing real-time segmentation approaches (such as the standard YOLO-seg model) typically employ fixed sampling rules (e.g. nearest neighbour or bilinear interpolation), limiting their flexibility and accuracy—particularly when handling the complex, intertwined boundary features of the axillary plexus. DFF-Net directly addresses this critical trade-off. By integrating a reparameterised feature fusion network with a SimBiF bidirectional feature pyramid network based on similarity attention modules, alongside an innovative DySample dynamic upsampling module, we achieve adaptive cross-scale feature fusion and dynamic learning offsets. This dual innovation precisely delineates intricate boundaries within neurovascular entanglements while preserving computational efficiency, transcending the inherent limitations of pure speed or pure accuracy approaches. This synergistic effect underscores our core methodological innovation: achieving high-precision instance segmentation at real-time speeds.

To further enhance segmentation performance, the innovatively optimised composite loss function combines dynamic weighted classification loss (SlideLoss-IoU), WIoU, and distribution focus loss (DFL), along with bounding box-guided mask cropping and loss normalisation, accelerating model training convergence and improving classification confidence and localisation accuracy. Experiments show that DFF-Net achieves 0.604 mAP@0.5:0.95 at a real-time inference speed of 57 FPS on a single GPU on the axillary brachial plexus dataset UABP, surpassing benchmark models such as Mask R-CNN and YOLOv8-seg, demonstrating excellent segmentation accuracy and computational efficiency. DFF-Net holds significant potential for application in ultrasound-guided nerve blockades. Its lightweight and high real-time characteristics make it easy to integrate into resource-constrained ultrasound devices, significantly improving the efficiency and safety of intraoperative navigation.

However, this study still has certain limitations that must be addressed in future work. First, the current model has limited robustness to low-quality ultrasound images and noise, and complex boundary segmentation still needs further optimisation, particularly in highly artifact-ridden scans. Future plans include incorporating self-supervised learning and contrastive learning strategies to enhance the model's generalisation ability against image degradation. Second, while DFF-Net performs exceptionally well on the axillary brachial plexus, its generalizability to other peripheral nerve regions, such as the cervical or interscalene levels, requires further validation and adaptation. These regions present distinct anatomical challenges, including different nerve bundle arrangements, varying tissue depths, and unique acoustic shadowing characteristics, which our model has not yet been explicitly trained to handle. Therefore, extending its application to these critical neural segmentation tasks will necessitate structure and algorithm optimisation to maintain efficiency and accuracy across diverse anatomical contexts. Additionally, DFF-Net's lightweight design and high real-time performance make it easy to integrate into ultrasound devices, enhancing the efficiency and safety of intraoperative navigation. It holds broad application potential in regional anaesthesia and neural imaging analysis and may become a cornerstone of ultrasound-guided technology.

Conclusion

This study explores the challenges faced in image segmentation of the axillary brachial plexus, aiming to effectively assist anaesthetists in performing peripheral nerve blocks. In axillary brachial plexus ultrasound image segmentation technology, challenges such as nerve-vessel intertwining, speckle noise, and structural blurring still exist. To address these issues, this study proposes a lightweight multi-instance segmentation model, DFF-Net, to provide efficient and precise technical support for ultrasound-guided brachial plexus block. Specifically, DFF-Net combines RepViT with spatial pyramid pooling to construct a lightweight multi-scale feature extraction architecture. It designs a bidirectional feature pyramid network (SimBiF) based on similarity attention modules and a dynamic upsampling module (DySample) for adaptive feature fusion. Additionally, it innovatively optimises the composite loss function to accelerate model training convergence and enhance classification confidence and localisation accuracy. The results demonstrate that the proposed method achieves superior performance in multi-object segmentation tasks compared to state-of-the-art methods, particularly in the segmentation of neural structures in axillary brachial plexus ultrasound images, marking a significant step toward deployable real-time medical imaging guidance.

Author Contribution

All authors contributed to the study conception and design. Material preparation, data collection and analysis were performed by Xinru Wu and Zhihao Yang. The first draft of the manuscript was written by Xinru Wu and all authors commented on previous versions of the manuscript. All authors read and approved the final manuscript.

Funding

This work was supported by the National Natural Science Foundation of China (62176031) and the Fundamental Research Funds for the Central Universities(2023CDJYGRHZD05).

Data Availability

The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.

Declarations

Ethics approval

We confirm that this research was conducted ethically in accordance with the World Medical Association's Code of Ethics (Declaration of Helsinki). The study protocol received formal approval from two institutional ethics committees: the Ethics Committee of Chongqing University Cancer Hospital (approval reference number: 2020 Lun (054), issued in 2020), and the Institutional Review Board (IRB) of Shenzhen Bao'an Traditional Chinese Medicine Hospital (approval reference number: KY-2025–110-01, issued on August 18, 2019). The study was carried out in full compliance with all relevant ethical regulations and guidelines.

Consent to participate

This article does not contain any studies involving human subjects.

Competing interests

The authors have no relevant financial or non-financial interests to disclose.