report2
report2
report2
PROJECT REPORT
OF MINOR PROJECT
BACHELOR OF TECHNOLOGY
Computer Science
SUBMITTED BY
Feasibility Study 12
Facilities required 14
Expected outcome 15
Conclusion and Future Scope 16
Reference 17
Introduction
Convolutional Neural Networks (CNNs), in particular, have rapidly advanced in deep learning
techniques, ushering in a new era of biological image analysis. Of all the uses for these
developments, biological image segmentation is one of the most important and essential
applications. For clinical diagnosis, treatment planning, and medical research, the capacity to
precisely define the boundaries of anatomical structures or diseased regions in medical imaging
is essential. Optimizing CNNs, which have proven to be exceptionally effective at identifying
complex patterns in data, is now essential to improving biological image segmentation's
precision and productivity.
In-depth investigation of Biomedical CNN-based
image segmentation is undertaken in this work, which explores the nuances of several CNN
architectures and a range of datasets. Our research aims to advance the state-of-the-art in the
crucial field of medical image analysis, building upon the groundbreaking work of scholars such
as Cui et al. (2019), who cleverly combined CNNs with multi-objective algorithms for malicious
code detection, and the ground-breaking U-Net model proposed by Ronneberger et al. (2015)
specifically designed for biomedical image segmentation.
Our research is motivated by the need to close the gap between complex difficulties in biomedical
imaging and state-of-the-art deep learning techniques developed in many areas. Medical images
are extremely complicated and variable due to a variety of modalities (such as MRI, CT, or X-
ray) and clinical circumstances, necessitating the use of robust and customized segmentation
techniques. Examining several CNN architectures—each with special advantages and
subtleties—becomes crucial in this situation. We aim at elucidating the subtleties of these
designs' performance in the context of biomedical picture segmentation through thorough
assessment and comparison.
Recent developments in adjacent disciplines, such as the work of Li et al. (2021) on multimodal
image fusion approaches and the studies of Abdar et al. (2021) on direct and cross-based binary
residual feature fusion for medical image classification, serve as motivation for us. Our research
technique incorporates the insights learnt from these studies to evaluate CNN models thoroughly
on a variety of biomedical datasets. We base our
evaluation system on these datasets, which have been carefully selected to capture the variety
prevalent in real-world clinical circumstances.
We aim to analyse the subtleties of their performance through methodical testing with CNN
architectures, from the more recent breakthroughs such as Swin Transformer (Liu et al., 2021) to
the basic models like U-Net. Furthermore, our comparative research will go beyond the
architectural domain and include datasets that encompass the wide range of biological imaging
problems. Our study aims to identify the best model-dataset combinations by carefully assessing
metrics like accuracy, sensitivity, specificity, and Dice coefficient. This will provide valuable
information for researchers and practitioners who are looking for exact solutions for particular
biomedical image segmentation tasks.
The upcoming sections of this research project will cover the complex techniques we used,
provide in-depth analyses of our experiments, communicate our findings, and have thought-
provoking conversations. By conducting this thorough investigation, we hope to not only make
a substantial contribution to the field of Biomedical CNN-based image segmentation but also
provide a nuanced understanding of the interactions between various CNN architectures and
datasets, advancing the field into new frontiers in terms of precision, effectiveness, and
practicality in real-world medical scenarios.
RequirementAnalysisandSystemSpecification
In the realm of malicious code detection, the studies by Cui et al. and Cui et al. present insights
into leveraging convolutional neural networks (CNNs) and multi-objective algorithms for robust
detection mechanisms. These findings are critical for establishing the requirements of efficient
and adaptive cybersecurity systems capable of identifying evolving threats. The research by
Wang et al extends the computational focus to high-performance computing for cyber-physical
social systems, emphasizing the need for evolutionary multi-objective optimization algorithms.
Such algorithms become integral requirements for ensuring the scalability and adaptability of
computing systems in dynamic socio-technical environments.
In the domain of image recognition and computer vision, the seminal works of Simonyan and
Zisserman He et al. and Howard et al. present pivotal advancements in deep convolutional
networks, residual learning, and efficient neural network architectures. These contributions set
forth requirements for the development of image recognition systems with heightened accuracy,
reduced computational complexity, and adaptability to diverse applications. The subsequent
papers by Liu et al. , Huang et al. , and Liu et al. introduce transformative approaches such as
global-attention-based networks, densely connected convolutional networks, and novel
convolutional architectures for the evolving landscape of computer vision, emphasizing the
ongoing demand for innovative methodologies.
Transitioning to medical image analysis, the works of Ronneberger et al. , Abdar et al. , and Li
et al. focus on segmentation and classification techniques using U-Net, binary residual feature
fusion, and multimodal image fusion. These contributions articulate the requirements for precise
and reliable medical image processing, catering to the intricate needs of healthcare applications.
Collectively, the analyzed references underscore the imperative for sophisticated algorithms,
adaptive optimization strategies, and innovative neural network architectures to meet the
evolving requirements of cybersecurity, computer vision, and medical imaging applications in
the contemporary technological landscape.
fig(b):-- Visual comparison of different input images and mask images of ISIC 2017 dataset using various
model
Net, these methods demonstrate the complexity of assessment metrics that are customized for
particular biomedical imaging problems.
Most importantly, the interaction between sophisticated designs (e.g., U-Net, ResUNet+) and
complex datasets is investigated. Ronneberger et aland Szegedy et al. have produced
The objectives derived from the provided references represent a comprehensive set of research
goals across various domains, each contributing to advancing our understanding and application
of cutting-edge technologies. These objectives encompass areas such as cybersecurity, computer
vision, optimization, and medical imaging, offering a roadmap for researchers and practitioners
to address critical challenges and explore opportunities for innovation.
In the realm of cybersecurity, the primary objective is to develop and evaluate advanced models
for malicious code detection. Reference 1 introduces the utilization of convolutional neural
networks (CNNs) and multi-objective algorithms in this context. By implementing these
techniques, the research aims to enhance the accuracy and efficiency of identifying malicious
code within software systems. In a world increasingly plagued by sophisticated cyber threats,
such as malware and viruses, the need for robust and real-time threat identification is paramount.
Reference 2 further complements this objective by focusing on the detection of malicious code
variants, demonstrating the continuous evolution of cybersecurity challenges. These objectives
are not only timely but also essential for safeguarding digital systems and data from ever-
evolving cyber threats.
Moving to the domain of computer vision, several objectives revolve around leveraging state-of-
the-art techniques for image analysis and understanding. Visual object tracking, as discussed in
Reference 8, represents a fundamental task in computer vision, with applications in surveillance,
autonomous vehicles, and augmented reality. The objective here is to explore advanced methods,
such as adaptive structural convolutional networks, to improve the accuracy and robustness of
object tracking, especially in challenging scenarios. This objective contributes to enhancing the
reliability and performance of computer vision systems in real-world applications.
Statistical comparison with different state-of the art methods on ISIC 2017 Dataset
MODELS ACCURACY(%) VALIDATION LOSS
ConUnext
94.84 5.16
93.83
U-NET 6.17
seminal papers that highlight the critical role these models play in CVC-ClinicDB, Retina MASK
SEGMENTATION, and Skin Lesion ISIC 2017, as well as their correctness and efficiency. In
addition, the review examines how flexible models such as AttU-Net may be in managing
intricate medical imaging details, demonstrating how they can advance the field of biomedical
image segmentation.
To sum up, this review carefully combines knowledge from several datasets and advanced
models, including U-Net, ResUNet+, ConvUNeXt, and AttU-Net. Through a comprehensive
synthesis of these results, the review acts as a scientific lighthouse, illuminating the complex
terrain of performance, customized assessment techniques, and subtle implementation
approaches in biomedical picture segmentation. The review contributes to the scholarly
conversation with its thorough analysis and offers invaluable direction for scholars navigating
the intricacies of deep learning in biomedical imaging.
Convolutional neural networks like UNet are made for applications like semantic picture
segmentation. It is made up of a symmetric expanding path that allows for exact localisation and
a contracting path that records context. The convolutional layers of a typical convolutional neural
network, which extract features and minimize spatial information, are comparable to the
contracting route. In contrast, the expanding path contains of convolutional layers that eventually
recover the spatial information to create a segmented image after up-sampling layers. Because of
its accurate segmentation capabilities, UNet's architecture finds widespread application in
medical image analysis tasks including cell segmentation and tumor identification.
Inspired by ResNet, ResUNet+ is an extension of the UNet design that uses residual connections
to solve the vanishing gradient issue and make deeper network training possible. Deep networks
can be trained because residual connections allow the gradient to pass through the network
directly without degrading. ResUNet+ enhances the model's capacity to capture both high-level
and low-level information by adding residual connections to the UNet architecture's contracting
and expanding channels. This design has proven very useful for jobs like satellite image analysis
and medical image segmentation, where handling complex structures and catching minute details
in the images are critical.
autonomous vehicles and surveillance systems, where image quality can significantly impact
decision-making.
In the field of medical image processing, the objectives revolve around improving the accuracy
and reliability of diagnostic systems. References 28 and 30 introduce approaches that leverage
convolutional neural networks and bilateral filters for medical image segmentation and fusion.
The research objectives aim to enhance the interpretation of medical images, potentially leading
to more accurate and timely diagnoses. Furthermore, Reference 29 introduces the concept of
uncertainty-aware modules, with the objective of developing diagnostic systems that not only
provide accurate results but also quantify the uncertainty associated with each diagnosis. This is
vital for medical practitioners and decision-makers, as it enables them to make more informed
and reliable decisions based on medical imaging data.
Attention UNet, or AttU-Net for short, is a UNet architecture that enhances its ability to capture
long-range dependencies and focus on pertinent areas of the input image by adding attention
methods. By enabling the network to selectively focus on informative portions of the input,
attention mechanisms facilitate more accurate feature extraction. By adding attention gates to the
UNet architecture's skip connections, AttU-Net allows the network to assess the significance of
characteristics at various scales. Spatial dependencies and complex patterns in the input data can
be efficiently captured by AttU-Net by dynamically varying the weights assigned to the features.
This design is especially helpful for tasks involving natural language processing, such as text
summarization and machine translation, and remote sensing and satellite image analysis, where
the link between distant pixels is critical
In this work, we make use of three different datasets, each carefully selected to reflect particular
difficulties in the field of biomedical picture segmentation. The first dataset is an extensive
collection of endoscopic images called CVC-ClinicDB. With 612 photos carefully divided into
training and testing sets, it provides a balanced representation for reliable analysis. Retina Mask
Segmentation, the second dataset, provides an equally distributed set of 1056 high-resolution
retinal pictures for training and testing. In conclusion, the Skin Lesion Segmentation ISIC 2017
dataset is a large collection that includes 2000 testing photos and 4000 training images that
Following our computation of accuracy and validation loss, we took the dataset and ran epochs
up to 20. Based on our results, we displayed the accuracy and validation loss graph, which is
shown below
effectively capture the intricacies of dermatological imaging. Our study is made possible by the
rich foundation these diverse datasets provide, which allows us to thoroughly explore and
evaluate different deep learning models across a range of biomedical imaging domains.
CVC-ClinicDB: There have been a total of 612 images used for training and 612 images are
used for testing.
Retina Mask Segmentation: There have been a total of 1056 images used for training and 1056
images are used for testing.
Skin Lesion Segmentation ISIC 2017: There have been a total of 4000 images used for training
and 2000 images are used for testing.
Literature Review
The literature review based on the provided references offers a glimpse into the dynamic and
multifaceted landscape of cutting-edge research across various domains, including cybersecurity,
computer vision, optimization, and medical imaging.
In the domain of cybersecurity, References 1 and 2 shed light on the critical issue of malicious
code detection. These references highlight the adoption of deep learning techniques, particularly
convolutional neural networks (CNNs) and multi-objective algorithms, to bolster the accuracy
and efficacy of identifying malicious code. Given the constantly evolving nature of cyber threats,
these approaches play a pivotal role in fortifying digital systems and enabling real-time threat
identification.
The realm of computer vision, as represented by References 8 to 27, is characterized by its
diversity and innovation. Visual object tracking, image recognition, and semantic segmentation
are at the forefront of research. Researchers are actively exploring advanced techniques, such as
adaptive structural convolutional networks and transformer architectures. These advancements
hold immense promise in improving tracking accuracy, image recognition, and pixel-level object
segmentation, thereby impacting applications in autonomous vehicles, medical image analysis,
and more.
Optimization and scheduling, discussed in References 4, 5, and 6, are essential in industrial and
operational contexts. These references introduce the application of differential evolution
algorithms and multi-objective optimization techniques to tackle complex scheduling challenges.
This research aims to enhance resource allocation and decision-making, a critical aspect in
manufacturing and cyber-physical systems
In the field of medical image processing, as outlined in References 28, 29, and 30, the focus is
on improving accuracy and reliability in diagnostic systems. Techniques such as U-net
architectures, uncertainty-aware modules, and multimodal fusion contribute to more precise
image segmentation and enhanced diagnostic interpretability. This research strives to provide
medical practitioners with diagnostic systems that not only deliver accurate results but also
quantify the associated uncertainty, thereby enabling more informed medical decisions.
In summary, these references collectively reflect the evolving state of technology and science
across diverse domains. Researchers and practitioners can draw inspiration from these findings
to drive their work forward, ultimately contributing to the advancement of these fields and
addressing real-world challenges
Feasibility Study
Technical Feasibility:
Data Availability and Quality: Evaluate the availability and quality of biomedical image datasets
suitable for segmentation tasks. Ensure that diverse datasets representing different medical
imaging modalities are accessible and that they cover a wide range of medical conditions.
Computational Resources: Assess the computational resources required for training and testing
convolutional neural network (CNN) models. Ensure that the hardware, such as GPUs, and
software tools for deep learning are available and can handle the computational demands of
training complex models.
Economic Feasibility:
Budget Analysis: Conduct a comprehensive budget analysis to estimate the costs associated with
data acquisition, hardware, software, personnel, and other project-related expenses.
Cost-Benefit Assessment: Evaluate the potential benefits of the project in terms of improved
medical image segmentation. Consider the impact on medical diagnosis, treatment planning, and
patient outcomes to justify the investment in the study.
Operational Feasibility:
Resource Availability: Ensure that the necessary human resources, including data scientists,
researchers, and medical experts, are available for the project. Define roles and responsibilities
within the team.
Project Timeline: Create a detailed project timeline with milestones and deliverables. Assess
the feasibility of adhering to the planned timeline, considering potential challenges and
delays.The methodology for the "Cross-Dataset Comparative Study on Biomedical CNN Models
for Image Segmentation" involves a systematic and rigorous approach to assess the performance
and generalization capabilities of convolutional neural network (CNN) models when applied to
diverse biomedical image datasets. First, a selection of representative and publicly accessible
biomedical image datasets is made, encompassing different imaging modalities and medical
conditions. Subsequently, the selected datasets undergo data preprocessing, which includes
image resizing, normalization, and augmentation to ensure uniformity and readiness for model
training. CNN models renowned for their efficacy in biomedical image segmentation, such as U-
Net, SegNet, and DeepLab, are chosen for evaluation. These models are trained on specific
training subsets from the datasets, while validation sets are employed to fine-tune
hyperparameters and assess model performance. The models' generalization capabilities are put
to the test by evaluating their segmentation accuracy, employing standard metrics like Dice
coefficient and IoU, on previously unseen data from testing sets within the various datasets.
Statistical analyses, such as ANOVA or t-tests, are performed to determine the presence of
significant performance variations across models and datasets, thereby identifying models that
exhibit consistent excellence or dependency on dataset characteristics. Interpretability and
visualization techniques are employed to gain insights into model behavior and the segmentation
results. The study culminates in a thorough discussion of findings and their implications, offering
recommendations for the use of CNN models in biomedical image segmentation and suggesting
avenues for further research, including model ensemble strategies and domain adaptation
techniques. A comprehensive research report or scientific paper is subsequently prepared to
transparently convey the methodology, results, discussions, and conclusions, facilitating
reproducibility and knowledge dissemination in the field of medical image analysis
Facilities Required
Implementing and comparing CNN-based semantic segmentation methods for dermoscopic
images in biomedical applications requires a set of specialized facilities. Firstly, a robust
computing infrastructure is essential. This should include high-performance workstations or
servers equipped with powerful GPUs to accelerate the training of deep learning models.
Additionally, access to parallel computing resources, such as a GPU cluster or cloud computing
platform, can significantly expedite experimentation and evaluation. Adequate storage capacity
is imperative for handling the large datasets typically involved in biomedical imaging. Moreover,
software tools are crucial for efficient model development and evaluation. This includes deep
learning frameworks like TensorFlow or PyTorch, as well as libraries for image processing and
visualization. In parallel, a comprehensive dataset of dermoscopic images is essential for training
and testing the models. Ensuring data privacy and compliance with ethical standards is
paramount, especially when working with medical imagery. Access to dermatology expertise is
highly beneficial, as it enables expert annotation of images for ground truth generation and
validation. Lastly, a structured experimental protocol, possibly integrating automated testing and
validation procedures, is vital for a systematic and unbiased comparison of different
segmentation methods. By leveraging these facilities, researchers can effectively implement,
evaluate, and compare CNN-based semantic segmentation methods for dermoscopic images,
advancing the field of biomedical image analysis.
Expected Outcomes
The implementation, performance, and comparison of CNN-based semantic segmentation
methods for dermoscopic images in biomedical applications are expected to yield significant
advancements in computer-aided diagnosis systems for dermatology. By leveraging state-of-the-
art techniques from the referenced studies, such as FFNet, Self-Attention Feature Fusion
Network, and Simple and Efficient Architectures, the anticipated outcomes include highly
accurate and efficient segmentation models. These models are poised to excel in scenarios with
limited annotated data, demonstrating robustness in few-shot learning settings. Additionally, the
incorporation of transformer-based architectures, as demonstrated by Li et al., is expected to
enhance the models' ability to capture complex spatial dependencies crucial in dermoscopic
image analysis. The FF-UNet, introduced by Iqbal et al., showcases potential for multimodal
biomedical image segmentation, indicating the adaptability and versatility of these models across
various imaging modalities. Integration of attention mechanisms and novel loss functions, as
demonstrated by Abraham and Khan, is anticipated to enhance the network's ability to discern
subtle features in dermoscopic images, ultimately improving segmentation accuracy. Moreover,
the exploration of end-to-end object detection with transformers, inspired by Carion et al., may
provide insights into capturing intricate relationships within dermoscopic images. These
advancements collectively aim to revolutionize dermatological diagnosis by automating the
process of lesion identification and analysis, ultimately contributing to more effective and timely
clinical interventions. Additionally, the proposed architectures, including FF-UNet, SIL-Net, and
Pyramid Residual Attention Module, promise to refine the accuracy of dermoscopic image
segmentation, potentially enabling more precise identification of skin lesions. These expected
outcomes not only have the potential to significantly impact the field of dermatology but also
hold promise for broader applications in biomedical image analysis and computer-aided
diagnosis systems.
References
1. Cui, Z., Du, L., Wang, P., Cai, X., & Zhang, W. (2019). Malicious code detection based on CNNs and
multi-objective algorithm. Journal of Parallel and Distributed Computing, 129, 50-58.
2 . Cui, Z., Xue, F., Cai, X., Cao, Y., Wang, G. G., & Chen, J. (2018). Detection of malicious code variants
based on deep learning. IEEE Transactions on Industrial Informatics, 14(7), 3187-3196.
3. Zhang, K., Su, Y., Guo, X., Qi, L., & Zhao, Z. (2020). MU-GAN: Facial attribute editing based on multi-
attention mechanism. IEEE/CAA Journal of Automatica Sinica, 8(9), 1614-1626.
4. Gao, D., Wang, G. G., & Pedrycz, W. (2020). Solving fuzzy job-shop scheduling problem using DE
algorithm improved by a selection mechanism. IEEE Transactions on Fuzzy Systems, 28(12), 3265-3275.
5. Wang, G. G., Cai, X., Cui, Z., Min, G., & Chen, J. (2017). High performance computing for cyber
physical social systems by using evolutionary multi-objective optimization algorithm. IEEE Transactions
on Emerging Topics in Computing, 8(1), 20-30.
6. Wang, G. G., Gao, D., & Pedrycz, W. (2022). Solving multiobjective fuzzy job-shop scheduling
problem by a hybrid adaptive differential evolution algorithm. IEEE Transactions on Industrial
Informatics, 18(12), 8519-8528.
7. Wang, G. G., & Tan, Y. (2017). Improving metaheuristic algorithms with information feedback
models. IEEE transactions on cybernetics, 49(2), 542-555.
8. Yuan, D., Li, X., He, Z., Liu, Q., & Lu, S. (2020). Visual object tracking with adaptive structural
convolutional network. Knowledge-Based Systems, 194, 105554.
9. Liu, P., Zhou, Y., Peng, D., & Wu, D. (2020). Global-attention-based neural networks for vision
language intelligence. IEEE/CAA Journal of Automatica Sinica, 8(7), 1243-1252.
10. Simonyan, K., & Zisserman, A. (2014). Very deep convolutional networks for large-scale image
recognition. arXiv preprint arXiv:1409.1556
11. Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., ... & Rabinovich, A. (2015). Going
deeper with convolutions. In Proceedings of the IEEE conference on computer vision and pattern
recognition (pp. 1-9).
12. Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., & Wojna, Z. (2016). Rethinking the inception
architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern
recognition (pp. 2818-2826).
13. He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep residual learning for image recognition.
In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 770-778).
14. He, K., Zhang, X., Ren, S., & Sun, J. (2016). Identity mappings in deep residual networks.
In Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October
11–14, 2016, Proceedings, Part IV 14 (pp. 630-645). Springer International Publishing.
15. Xie, S., Girshick, R., Dollár, P., Tu, Z., & He, K. (2017). Aggregated residual transformations for
deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern
recognition (pp. 1492-1500).
16. Howard, A. G., Zhu, M., Chen, B., Kalenichenko, D., Wang, W., Weyand, T., ... & Adam, H. (2017).
Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint
arXiv:1704.04861
17. Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., & Chen, L. C. (2018). Mobilenetv2: Inverted
residuals and linear bottlenecks. In Proceedings of the IEEE conference on computer vision and pattern
recognition (pp. 4510-4520).
18. Huang, G., Liu, Z., Van Der Maaten, L., & Weinberger, K. Q. (2017). Densely connected
convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern
recognition (pp. 4700-4708).
19. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., ... & Houlsby,
N. (2020). An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint
arXiv:2010.11929.
20. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., ... & Guo, B. (2021). Swin transformer:
Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international
conference on computer vision (pp. 10012-10022).
21. Liu, Z., Mao, H., Wu, C. Y., Feichtenhofer, C., Darrell, T., & Xie, S. (2022). A convnet for the 2020s.
In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (pp. 11976-
11986).
22. Long, J., Shelhamer, E., & Darrell, T. (2015). Fully convolutional networks for semantic
segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp.
3431-3440).
23. Zhao, H., Shi, J., Qi, X., Wang, X., & Jia, J. (2017). Pyramid scene parsing network. In Proceedings
of the IEEE conference on computer vision and pattern recognition (pp. 2881-2890).
24. Chen, L. C., Papandreou, G., Kokkinos, I., Murphy, K., & Yuille, A. L. (2014). Semantic image
segmentation with deep convolutional nets and fully connected crfs. arXiv preprint arXiv:1412.7062.
25. Chen, L. C., Papandreou, G., Kokkinos, I., Murphy, K., & Yuille, A. L. (2017). Deeplab: Semantic
image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE
transactions on pattern analysis and machine intelligence, 40(4), 834-848.
26. Liu, K., Ye, Z., Guo, H., Cao, D., Chen, L., & Wang, F. Y. (2021). FISS GAN: A generative
adversarial network for foggy image semantic segmentation. IEEE/CAA Journal of Automatica
Sinica, 8(8), 1428-1439.
27. Liu, K., Ye, Z., Guo, H., Cao, D., Chen, L., & Wang, F. Y. (2021). FISS GAN: A generative
adversarial network for foggy image semantic segmentation. IEEE/CAA Journal of Automatica
Sinica, 8(8), 1428-1439.
28. Ronneberger, O., Fischer, P., & Brox, T. (2015). U-net: Convolutional networks for biomedical image
segmentation. In Medical Image Computing and Computer-Assisted Intervention–MICCAI 2015: 18th
International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18 (pp. 234-241).
Springer International Publishing.
29. Abdar, M., Fahami, M. A., Chakrabarti, S., Khosravi, A., Pławiak, P., Acharya, U. R., ... &
Nahavandi, S. (2021). BARF: A new direct and cross-based binary residual feature fusion with
uncertainty-aware module for medical image classification. Information Sciences, 577, 353-378.
30. Li, X., Zhou, F., Tan, H., Zhang, W., & Zhao, C. (2021). Multimodal medical image fusion based on
joint bilateral filter and local gradient energy. Information Sciences, 569, 302-325.