Research on Automatic Medical Image Segmentation and Boundary Optimization Algorithm

Based on Deep Feature Learning

Abstract: To address the challenges of blurred boundaries, complex structures, and significant
scale differences in medical image segmentation, this paper proposes an automatic medical image
segmentation and boundary optimization algorithm based on deep feature learning. This method,
building upon an encoder-decoder segmentation framework, introduces multi-scale feature fusion
and channel semantic enhancement mechanisms to improve the model's ability to represent
complex structures. Furthermore, it achieves fine-grained optimization of the segmentation
contour through boundary-aware constraints and a joint loss function. In multiple comparative
experiments, the proposed method outperforms traditional segmentation models in terms of Dice,
IoU, and ASSD, achieving a Dice of 0.903, an IoU of 0.824, and an ASSD of 1.42, demonstrating
higher region overlap accuracy and boundary localization accuracy. Ablation experiments further
validate the effectiveness of the multi-scale feature fusion and boundary constraint strategies in
improving segmentation performance. The results indicate that the proposed method can
significantly improve boundary quality while maintaining overall segmentation accuracy,
providing an effective solution for intelligent medical image analysis.
Keywords: Medical image segmentation; Deep feature learning; Multi-scale feature fusion;
Boundary optimization; Joint loss function
1. Introduction
With the continuous development of medical imaging technology, multimodal data such as
CT, MRI, ultrasound and pathological images play an increasingly important role in clinical
diagnosis, treatment planning and efficacy evaluation. Medical images contain rich structural and
functional information, and accurate segmentation of target organs or lesion areas is the basic link
to achieve quantitative analysis and intelligent auxiliary decision-making. However, medical
image segmentation tasks face multiple challenges such as complex tissue structure, blurred
boundaries, uneven gray distribution and serious noise interference. Traditional manual
segmentation methods are not only time-consuming and labor-intensive, but also highly
subjective, making it difficult to meet the needs of large-scale clinical applications [1]. In recent
years, the rapid development of deep learning technology has provided a new solution for
automatic medical image segmentation [2]. By constructing a multi-layer nonlinear feature
extraction model, deep neural networks can automatically learn high-level semantic information
from the original images, showing significant advantages in complex structure recognition.
Segmentation methods based on deep feature learning have achieved relatively ideal results in
multiple medical imaging tasks, but there are still some shortcomings in practical applications. On
the one hand, while focusing on the overall semantics of the region, the model often ignores fine-
grained boundary information, leading to discontinuities or excessive smoothing at the contours in
the segmentation results. On the other hand, the differences in feature representation between
structures of different scales make the model less stable when dealing with small targets or
complex-shaped regions.
To address the above problems, how to further improve the accuracy of boundary localization
and structural consistency while ensuring the overall segmentation accuracy has become a key
issue in current medical image segmentation research. Therefore, it is necessary to introduce a
boundary-aware mechanism and structural constraint strategy within the framework of deep
feature learning to effectively optimize the segmentation results. Based on this, this paper focuses
on the task of automatic medical image segmentation, studies a segmentation algorithm that
integrates deep feature learning and boundary optimization, constructs a segmentation model that
takes into account both global semantics and local boundary information, and verifies its
effectiveness in terms of segmentation accuracy and boundary quality through experiments,
providing more reliable technical support for intelligent medical image analysis.
2. Literature Review
Medical image segmentation, as an important research direction in the field of medical image
analysis, has long received widespread attention. Early research mainly relied on rule-based and
model-driven methods, achieving target region segmentation through artificial modeling of image
grayscale, texture, and spatial information. These methods typically employ threshold
segmentation, region growing, edge detection, and energy function-based optimization strategies,
achieving certain results in images with simple structures and high contrast. However, due to the
prevalence of noise interference, variable organ morphology, and blurred tissue boundaries in
medical images, traditional methods are less robust in complex scenarios and struggle to meet
diverse clinical needs [3-4]. With the development of machine learning technology, some studies
have begun to introduce statistical learning models, improving segmentation performance through
a combination of feature engineering and classifiers. These methods alleviate the limitations of
manual rule design to some extent, but their effectiveness is highly dependent on the quality of
feature selection and it is difficult to automatically extract discriminative semantic features from
high-dimensional data. Against this backdrop, deep learning methods have gradually become the
mainstream direction of medical image segmentation research [5]. Segmentation models based on
deep neural networks can achieve direct mapping from the original image to the segmentation
result through end-to-end training, effectively reducing manual intervention and improving
segmentation efficiency and consistency.
Existing deep learning segmentation methods mostly adopt an encoder-decoder structure,
extracting high-level semantic features through layer-by-layer downsampling and then restoring
spatial resolution by combining upsampling [6]. This structure has a clear advantage in capturing
global context information, enabling the model to accurately locate the target region. However,
due to the inevitable loss of spatial information introduced by the downsampling process, the
model often has shortcomings in recovering fine boundaries, especially in areas where the contrast
between the target edge and the background is low, and the segmentation results are prone to
contour shift or boundary unsmoothness. To alleviate this phenomenon, some studies have tried to
introduce multi-scale feature fusion strategies to enhance the model's ability to express detailed
structures by integrating feature information at different levels. On this basis, in response to the
problem of blurred boundaries and complex structures in medical images, more and more studies
have begun to focus on the role of boundary information in segmentation tasks. Related methods
usually guide the model to pay more attention to the target contour region during training by
introducing edge detection branches, boundary attention mechanisms or structural constraints[7].
These methods improve the boundary localization accuracy to a certain extent, but their effect
often depends on the quality of boundary feature extraction and the design of constraint methods.
If not handled properly, it may lead to unstable model training or excessive response to noisy
edges.
In addition, the design of the loss function is also an important factor affecting the
performance of medical image segmentation[8]. Traditional pixel-level loss functions mainly
focus on the degree of region overlap, have low sensitivity to boundary errors, and are difficult to
effectively constrain the morphological consistency of the segmented contour. Therefore, some
studies have proposed combining region loss and boundary-related loss to simultaneously improve
region accuracy and boundary quality through multi-objective optimization. These methods have
achieved some success in improving boundary continuity, but there is still room for improvement
in complex structures or small target scenarios.
Overall, existing research has made significant progress in automatic medical image
segmentation, but it still generally faces the challenge of balancing deep feature representation
with fine boundary modeling. On the one hand, relying solely on high-level semantic features
makes it difficult to accurately depict detailed boundaries; on the other hand, overemphasizing
local edge information may weaken the model's understanding of the overall structure. Therefore,
how to effectively integrate global semantic information and local boundary constraints within a
deep feature learning framework to construct a medical image segmentation model that combines
accuracy and stability remains an important direction worthy of in-depth research. This paper,
based on a systematic analysis of existing research, focuses on deep feature learning and boundary
optimization strategies, striving to provide a more robust and efficient solution for automatic
medical image segmentation.
3. Research Methods
3.1 Overall Framework of Deep Feature Learning-Driven Medical Image Segmentation
The core objective of medical image segmentation is to accurately distinguish target regions
from background regions from high-dimensional, noisy, and structurally complex image data.
However, different organs or lesions exhibit significant differences in morphology, scale, and
texture distribution, making it difficult for single-scale or shallow features to simultaneously
depict global semantic information and local structural details. Therefore, in segmentation model
design, the first problem to be solved is how to construct a deep feature learning framework that
can simultaneously capture global context and local discriminative features.
 Based on the above considerations, this paper adopts an encoder-decoder network structure as
the overall framework of the segmentation model. The encoder extracts high-level semantic
features step-by-step through multi-layer convolution and downsampling operations to enhance
the model's perception of the overall target structure; the decoder restores spatial resolution
through layer-by-layer upsampling, gradually approximating the original image scale through
feature mapping. In this process, to avoid the loss of detail information caused by downsampling,
the model introduces feature skip connections between the encoding and decoding stages, fusing
shallow spatial detail information with deep semantic features to achieve collaborative expression
of multi-level features.
Let the input medical image be I ∈ R H × W × C, where H, W and C represent the spatial size and
number of channels of the image, respectively. The feature representation of the l-th layer in the
encoding stage can be defined as:
F l = σ ( W l * Fl -1 + b l ) (1)
Where, ∗ represents the convolution operation, Wl and bl are the convolution kernel
parameters and bias terms of the l-th layer, respectively, and σ( ⋅) is a non-linear activation
function. As the network depth increases, the receptive field of the feature map gradually expands,
enabling the model to learn more semantic region representations. In the decoding stage, to
achieve spatial reconstruction of features, an upsampling operator is introduced to restore the size
of the feature map, and it is concatenated with the corresponding coding layer features to obtain
the fused feature representation:
F^ l =Concat ( Up ( Fl +1 ), Fl ) (2)
This fusion method not only preserves high-level semantic information, but also enhances the
model's ability to perceive edge and detail regions, providing a foundation for subsequent fine
segmentation.
3.2 Multi-scale deep feature fusion and semantic enhancement modeling
Although the encoder-decoder structure can alleviate the problem of spatial information loss
to some extent, medical images often contain both large-scale organ structures and small-scale
lesion regions. Single-level features are still difficult to fully represent targets of different scales.
Therefore, it is necessary to introduce a multi-scale feature fusion mechanism into the network to
enhance the model's adaptability to scale changes.
In the feature fusion process, this paper extracts contextual information at different scales
through parallel convolution operations and aggregates it in the channel dimension to construct a
multi-scale semantic enhancement module. Let the input feature of a certain layer be F, and the
multi-scale feature representation is obtained through convolution kernels of different scales
F (k ) = σ ( W (k ) * F ) , k =1,2,…, K (3)
where different k correspond to different receptive field sizes, used to capture multi-level
structural information. Subsequently, the features of each scale are weighted and fused to obtain
the enhanced feature representation:
K
F m s = ∑ α k F( k ) (4)
k =1
Where αk is a learnable scale weight parameter used to adaptively adjust the importance of
features at different scales. To further enhance the model's ability to focus on key regions, this
paper introduces a channel attention mechanism based on multi-scale fusion. By adaptively
weighting feature channels, it highlights discriminative information related to target segmentation.
The channel attention weight can be expressed as:
w =Sigmoid ( W 2 δ ( W 1 GAP( F ms ) ) ) (5)
Where, GAP(⋅) represents the global average pooling operation, and δ( ⋅) is a non-linear
activation function. The final output feature is:
F en h = w⨀F ms (6)
Through the above process, the model effectively enhances its ability to express key
structures and fine-grained regions while maintaining global semantic consistency.
3.3 Boundary-aware constraints and joint loss optimization strategy
In medical image segmentation tasks, the accuracy of region overlap cannot fully reflect the
segmentation quality, especially in areas with blurred boundaries or tissue adhesions, where
boundary offset can significantly affect subsequent quantitative analysis. Therefore, this paper
introduces a boundary-aware constraint mechanism based on deep feature learning to explicitly
model the segmentation boundary from the perspective of optimization objectives.
First, by applying a spatial gradient operator to the predicted segmentation results, its
boundary response features are extracted.
Bp =|∇ P| (7)
where P is the probability segmentation map output by the model. Correspondingly, the
boundary representation of the true label is
Bg =|∇G | (8)
By minimizing the difference between the predicted boundary and the true boundary, the
model is guided to pay more attention to the contour region during training. In terms of loss
function design, this paper adopts a joint optimization strategy that combines region loss and
boundary loss. Region segmentation loss is defined as
2| P ∩G |
Ls eg =1- (9)
| P |+|G |
Boundary constraint loss is defined as
L bd =‖ B p - Bg ‖ 1 (10)
The final joint loss function is expressed as
L = Lseg + λ Lb d (11)
Where λ is the boundary constraint weight, used to balance the relationship between region
accuracy and boundary accuracy. During model training, the joint loss is optimized through
backpropagation mechanism, so that the model can simultaneously improve boundary continuity
and structural consistency while ensuring the overall segmentation accuracy. The parameters are
shown in Table 1 [9].
Table 1 Key model and training parameter settings
Parameter category Parameter name Numerical range
Number of encoding
4
layers
Network structure
Initial number of
64
channels
Number of
Multi-scale module 3
convolution scales K
Learning rate 1e-4
Training strategy
Batch size 8
Loss weight Boundary weight λ 0.3–0.6
Regularization Weight decay 1e-5
4. Research results
4.1 Comparative analysis of the overall performance of different segmentation models
To verify the effectiveness of the proposed automatic medical image segmentation method
based on deep feature learning and boundary optimization in overall segmentation performance, a
variety of typical segmentation models are selected as comparison methods, and performance
evaluation is carried out under the same dataset and unified experimental settings. Evaluation
metrics include Dice coefficient, IoU index, and average symmetric surface distance (ASSD),
which comprehensively measure model performance from the perspectives of regional overlap
and boundary deviation. Quantitative experimental results for each model are shown in Table 2.
Table 2 Comparative results of the overall performance of different segmentation models
Method category Dice IoU ASSD
Traditional encoder-
0.842 0.735 2.41
decoder model
Multi-scale feature
0.865 0.764 2.03
fusion model
Introducing the
Attention Mechanism 0.879 0.783 1.86
Model
Boundary Constraint 0.887 0.795 1.74
 Segmentation Model
Our Method 0.903 0.824 1.42
As can be seen from Table 2, there are significant differences in the segmentation
performance of different models. The traditional encoder-decoder model has a certain ability in
overall region localization, but due to insufficient characterization of fine-grained structures, its
Dice and IoU indices are relatively low, and it also shows a large boundary deviation in the ASSD
index, indicating that this method has certain limitations in contour localization. With the
introduction of the multi-scale feature fusion mechanism, the model's adaptability to structures of
different scales is enhanced, the region overlap index is improved, and the boundary distance
index also shows a significant decrease. After further introducing the attention mechanism, the
model can pay more attention to the key feature regions related to the segmentation target, so that
the interference of irrelevant background information is suppressed, and the overall segmentation
accuracy continues to improve. In contrast, the segmentation model with boundary constraints has
a limited improvement in region accuracy, but it has achieved a more significant improvement in
the ASSD index, indicating that boundary-related information has a positive effect on improving
contour localization accuracy. The proposed method achieves optimal results across all three
metrics, with significantly higher Dice and IoU scores compared to the comparative methods,
while the ASSD value is significantly reduced. This indicates that the proposed method not only
accurately segments the target region but also effectively reduces the deviation between the
segmentation boundary and the true contour. Overall, the proposed method achieves a relatively
ideal balance between region accuracy and boundary quality, validating the effectiveness of the
joint modeling strategy of deep feature learning and boundary optimization.
4.2 Analysis of the Effect of Boundary Optimization Strategy on Segmentation Accuracy
Improvement
To further verify the role of the boundary optimization strategy in the segmentation task, this
paper conducts a comparative experiment on whether or not boundary-aware constraints are
introduced while maintaining the consistency of the main network structure. By analyzing the
changes in model performance under different boundary constraint weight settings, the influence
of the boundary optimization mechanism on the segmentation results is evaluated. The
experimental results are shown in Table 3.
Table 3 Comparison of Segmentation Performance under Different Boundary Constraint Settings
Boundary constraint
Dice IoU ASSD
weight λ
0 (No Boundary
0.881 0.792 1.89
Constraint)
0.2 0.892 0.806 1.63
0.4 0.901 0.821 1.44
0.6 0.903 0.824 1.42
0.8 0.897 0.816 1.51
As shown in Table 3, without boundary constraints, the model mainly relies on region loss for
optimization, achieving relatively stable region overlap accuracy, but its performance on the
ASSD metric is relatively weak, indicating a certain offset between the segmentation boundary
and the true contour. As the boundary constraint weights gradually increase, the model's focus on
the contour region during training continuously strengthens, and the ASSD value shows a
significant downward trend, indicating a continuous improvement in boundary localization
accuracy. When the boundary weights are set in a moderate range, the model achieves a better
balance between region accuracy and boundary accuracy, with Dice and IoU scores increasing
simultaneously. This demonstrates that appropriate boundary constraints not only do not impair
the region segmentation effect but can also improve the overall segmentation quality through
structural constraints. However, when the boundary weights are further increased, the regional
accuracy index decreases slightly, while the ASSD value rebounds. This indicates that excessively
strong boundary constraints may cause the model to overemphasize local contour information,
thus affecting the accurate modeling of the overall regional structure. The above results show that
boundary-aware constraints have a significant positive effect in medical image segmentation, but
their weight settings need to be reasonably adjusted. The method in this paper establishes a
balance between regional and boundary information through a joint loss function, enabling the
model to maintain stable and high-quality segmentation performance even in complex structural
scenarios.
4.3 Ablation Experiment Analysis of Multi-Scale Feature Fusion Mechanism
To verify the contribution of the multi-scale deep feature fusion module to the improvement
of model performance, this paper designs multiple ablation experiments. By gradually introducing
features of different scales and semantic enhancement mechanisms, the changes in segmentation
performance are analyzed. The relevant experimental settings and results are shown in Table 4.
Table 4 Experimental Results of Multi-Scale Feature Fusion Ablation
Experimental
Dice IoU ASSD
Configuration
Single-Scale Feature 0.864 0.758 2.12
Dual-Scale Feature
0.883 0.789 1.76
Fusion
Three-Scale Feature
0.895 0.812 1.53
Fusion
Three-Scale +
Channel 0.903 0.824 1.42
Enhancement
As can be seen from Table 4, the single-scale feature model exhibits certain limitations when
dealing with medical images with significant scale differences, showing relatively low Dice and
IoU indices and large boundary errors. After introducing dual-scale feature fusion, the model can
simultaneously focus on local structure and medium-scale contextual information, significantly
improving segmentation accuracy and reducing the ASSD index. With a further increase in the
number of feature scales, the model's ability to represent complex structures and small targets
continues to improve, and both regional overlap and boundary indices show a steady improvement
trend. Especially after introducing a channel semantic enhancement mechanism, the model can
adaptively emphasize feature channels that contribute significantly to the segmentation task,
further reducing interference from irrelevant information, thus achieving the best results in all
three evaluation metrics. These ablation experiments demonstrate that multi-scale feature fusion is
a crucial factor in improving medical image segmentation performance, and combining it with
semantic enhancement strategies helps to further explore the discriminative power of deep
features. The above results experimentally verify the rationality and effectiveness of the multi-
scale deep feature learning module in our proposed method.
5. Conclusion
This paper focuses on the core problem of "difficulty in balancing regional accuracy and
boundary quality" in automatic medical image segmentation, and proposes a segmentation
algorithm that integrates deep feature learning and boundary optimization. The paper also
conducts a systematic study on model structure design, feature modeling methods, and
optimization strategies. By constructing an encoder-decoder segmentation framework and
introducing multi-scale feature fusion and channel semantic enhancement mechanisms, the model
effectively improves its ability to express complex structures and small targets while maintaining
global semantic consistency. Experimental results show that compared with the traditional
encoder-decoder model, the proposed method has significant advantages in overall segmentation
performance. The Dice and IoU indices are improved to 0.903 and 0.824, respectively, while the
ASSD is reduced to 1.42, indicating significant improvements in both region overlap accuracy and
boundary localization accuracy. Further experiments on boundary constraint weights show that
appropriately introducing boundary-aware constraints can effectively reduce boundary bias. When
the boundary weights are set within a reasonable range, the model achieves a good balance
between region accuracy and boundary accuracy. Ablation experiments involving multi-scale
feature fusion also verify the crucial role of collaborative modeling of features at different scales
in improving segmentation performance. In summary, the proposed segmentation algorithm
exhibits good robustness and stability in complex medical imaging scenarios, significantly
improving boundary continuity and structural consistency while maintaining overall segmentation
accuracy. Future research could further combine multimodal image information with lightweight
network structure design to enhance the model's generalization ability and computational
efficiency in practical clinical applications.
References
[1] An, F. P., & Liu, J. E. (2021). Medical image segmentation algorithm based on multilayer
boundary perception-self attention deep learning model. Multimedia Tools and Applications,
80(10), 15017-15039.
[2] Liu, X., Yang, L., Chen, J., Yu, S., & Li, K. (2022).Region-to-boundary deep learning model
with multi-scale feature fusion for medical image segmentation. Biomedical Signal
Processing and Control, 71, one hundred and three thousand one hundred and sixty-five
[3] Kim, M., & Lee, B. D. (2021). A simple generic method for effective boundary extraction in
medical image segmentation. IEEE Access, 9, 103875-103884.
[4] Inik, Ö., & Ülker, E. (2022). Optimization of deep learning based segmentation method. Soft
Computing, 26(7), 3329-3344.
[5] Fan, J., Liu, L., & Yu, X. (2025). BFLE-Net: Boundary Feature Learning and Enhancement
Network for Medical Image Segmentation. Electronics, 14(15), 3054.
[6] Zhang, X., Wang, D., & Chen, H. (2019). Improved biogeography-based optimization
algorithm and its application to clustering optimization and medical image
segmentation. IEEE Access, 7, 28810-28825.
[7] Mortazi, A., Cicek, V., Keles, E., & Bagci, U. (2023).Selecting the best optimizers for deep
learning–based medical image segmentation. Frontiers in Radiology, 3, one million one
hundred and seventy-five thousand four hundred and seventy-three
[8] Zhi, Z., & Qing, M. (2021). Intelligent medical image feature extraction method based on
improved deep learning. Technology and Health Care, 29(2), 363-379.
[9] Liao, W. X., He, P., Hao, J., Wang, X. Y., Yang, R. L., An, D., & Cui, L. G. (2019). Automatic
identification of breast ultrasound image based on supervised block-based region
segmentation algorithm and features combination migration deep learning model. IEEE
journal of biomedical and health informatics, 24(4), 984-993.