A Futurist Approach to Wards Image Enhancement Technologies in Contemporary Medical Research

A Futurist Approach to Wards Image Enhancement Technologies in Contemporary Medical Research

Gurpreet Kaur, Purnima Associate Professor, Assistant Professor P G Department of Computer Science,

SGGS College, affiliated to Panjab University, Chandigarh, India

Abstract: Medical imaging is pivotal in modern healthcare that influences diagnostic precision and treatment efficacy. Although, the challenges like low contrast, noise and artifacts frequently compromise image quality. This article presents a comprehensive overview of image enhancement techniques in medical imaging across classical methods to progressive advancements. We look into traditional techniques such as histogram equalization, spatial filtering and wavelet transforms and analyze their applications and limitations. The focus then shifts to the transformative impact of deep learning in medical image enhancement that includes convolutional neural networks and their applications. The article emphasizes the critical role of image enhancement in various medical imaging modalities including X-ray, computed tomography (CT), magnetic resonance imaging (MRI), ultrasound and positron emission tomography (PET). Furthermore, it discusses the evolving landscape of image enhancement research and addresses the challenges and proposing future directions. This article aims to provide a holistic understanding of image enhancement in medical imaging that fosters informed decision-making for researchers, practitioners and healthcare professionals.

IndexTerms – Medical Imaging, Image Enhancement, Traditional Methods, Machine Learning Methods, Healthcare.

I INTRODUCTION

Image enhancement is an essential task in image processing that aims to improve images by highlighting specific features or adding information through various methods. The main goals of image enhancement are to enhance differences between objects in images, suppress unimportant features, increase information content, and improve image quality, interpretation and recognition [1-2]. The various medical imaging technologies such as X-ray, computed tomography (CT), magnetic resonance imaging (MRI), ultrasound, and positron emission tomography (PET) generate images that form the basis of modern healthcare [3]. Though, these images often suffer from various inherent limitations such as low contrast, noise and artifacts. Therefore, the image enhancement techniques came into picture that aim to overcome these limitations and improve the quality of medical images which ultimately aiding healthcare professionals in making accurate diagnoses and treatment decisions [4]. This paper presents a comprehensive literature review of image enhancement algorithms. While there are existing reviews on this topic, they often focus on specific techniques or applications that neglect the basic differences and connections between methods. This paper fills this gap by providing a thorough review of both traditional and machine learning-based image enhancement methods. The traditional image enhancement algorithms are discussed that involve designing filtering techniques and modifying pixel values to improve image quality. The paper categorizes traditional enhancement methods into these four types named as gray level transformation methods, histogram equalization methods, retinex-based methods and filtering methods. Machine learning-based algorithms; on the other hand enhance low-light images by learning features from well-lit images with a large training dataset. The paper discusses four categories of machine learning-based methods which are convolutional neural networks (CNNs), generative adversarial networks (GANs), denoising auto encoders and image dehazing models and this categorization is shown in Figure 1given below.

Gray level Transformation

Histogram Equalization

Traditional Methods

Filtering Methods

Retinex Methods

Image Enhancement

Techniques

Convolutional Neural Networks

Machine Learning Methods

General Adversal Networks

Image Dehazing Models

Figure 1. Various Image Enhancement Techniques

To evaluate the performance of various image enhancement algorithms, the paper reproduces the traditional and machine learning-based methods and compares on different basis. The paper's structure includes an introduction to traditional image enhancement method, a discussion of traditional methods, machine learning-based methods, comparison of both the methods and concluding remarks and future research directions. Thus, this paper provides valuable insights into image enhancement techniques and their applications which will benefit researchers and readers seeking a deeper understanding of image processing algorithms.

II TRADITIONAL APPROACHES FOR IMAGE ENHANCEMENT

Traditional image enhancement methods are foundational techniques that have been widely used for decades to improve the visual quality of images. These methods often involve simple mathematical operations applied to pixel values or spatial filtering [5]. Some most popular traditional image enhancement methods have been discussed here that explore their principles, applications and limitations. Gray level transformation methods are fundamental for adjusting the contrast and brightness of an image. These methods operate on the pixel intensities and include linear transformations, logarithmic transformations, and power-law (gamma) transformations. Linear Transformation involves scaling and translation of pixel values to enhance contrast and adjust brightness. It's a simple yet effective method commonly used in image pre-processing. Logarithmic Transformation is particularly useful for expanding intensity levels in the lower range and also enhances details in darker regions of an image as well as reveals information that is obscured. Power-Law (Gamma) Transformation method allows for flexible adjustment of contrast by employing a power-law function. It is especially valuable in scenarios where a non-linear relationship between pixel values and perceived intensity is desired. Histogram Equalization is a widely used technique for contrast enhancement. It operates on the histogram of the image by redistributing pixel intensities to achieve a more uniform distribution. This helps in revealing details and improving the overall contrast of an image. This technique finds applications in medical imaging, satellite imaging and scenarios where images suffer from low contrast or limited dynamic range. However, histogram equalization lead to over-enhancement and at times causes unnatural-looking results. It also tends to be sensitive to outliers in the image that affects its performance. Retinex Method was proposed by Edwin H. Land and stands as a widely employed image enhancement technique grounded in scientific experimentation and analysis. This term was coined from the fusion of "retina" and "cortex," thus "Retinex" encapsulates a model based on three fundamental assumptions [6-7]:

1. The natural world lacks inherent color and the colors we perceive result from the interplay of light with objects. Considering an example, the water in human eyes is colorless, but a water-soap film appears colorful due to light interference on its surface.

2. Each color region is constituted by the primary colors- red, green and blue at specific wavelengths.

3. The color of each unit region is determined by these three primary colors.

The Retinex theory speculates that an object's color is defined by its capacity to reflect long-wave (red), medium-wave (green) and short-wave (blue) illumination, rather than the absolute value of the reflected light. Apart from various traditional linear and nonlinear methods, retinex method excels in achieving a harmonious balance in dynamic range compression, edge enhancement and color invariance. Over the years, the retinex algorithm has evolved from the single-scale retinex algorithm (SSR) to the multi-scale Retinex algorithm (MSR) and further to the multi-scale retinex algorithm with color restoration (MSRCR). These advancements emphasize its continued refinement and applicability across diverse image enhancement scenarios. Filtering Methods include two types of filtering techniques which are spatial and frequency domain filtering [8]. Spatial filtering involves the application of convolution masks to enhance or suppress certain features in an image. It has smoothing filters like the Gaussian filter that is employed to reduce noise and blur images and also making them useful for applications like image preprocessing. The other category of spatial filter is sharpening filters like the Laplacian that is used to enhance edges and details, making them crucial in scenarios where edge detection and feature enhancement are priorities. Frequency domain filtering involves transforming images into the frequency domain using techniques like the Fourier transform. In this domain, high-pass and low-pass filters can be applied for specific image enhancements. High-pass filtering enhances high-frequency components by emphasizing fine details and edges that is useful for sharpening images. Low-pass filtering reduces noise and smoothes images by suppressing high-frequency components. This is particularly valuable for image preprocessing to remove unwanted details. Median filtering is a popular method for noise reduction, especially in scenarios where images are affected by salt-and-pepper noise. Here, the pixel value of each point in the image is replaced with the median value of the pixel values in its local neighborhood. This type of filtering technique is widely used in medical imaging and remote sensing where preserving edges and fine details is crucial. There are few other conventional strategies namely, contrast stretching which is also known as intensity scaling. It is a straightforward method to enhance the overall contrast of an image and involves linearly scaling pixel intensities to span a wider range by improving the visual appearance of the image. This strategy is commonly used in real-time applications and scenarios where a quick enhancement of contrast is required. The other one is pseudo coloring that involves assigning colors to different intensity levels in an image. It enhances the visual discrimination of different intensities by representing them with distinct colors and often used in medical imaging for highlighting specific structures or anomalies in an image.

Traditional image enhancement methods have been the backbone of image processing for years. Each method addresses specific challenges that provide a range of tools for enhancing images based on different criteria. While these methods are foundational, they do have limitations especially when dealing with complex scenarios such as non-uniform illumination or varied structures within an image. In recent times, the advent of advanced techniques based on machine learning and deep learning has provided new avenues for overcoming these limitations and further improving the state of image enhancement. However, the simplicity, transparency and computational efficiency of traditional methods continue to make them valuable in various applications as well as ensure a holistic and diverse toolbox for image enhancement.

III MACHINE LEARNING METHODS OF IMAGE ENHANCEMENT

Machine learning (ML) has revolutionized image processing as it provides powerful tools for enhancing images in ways that traditional methods regularly struggle with. These ML techniques enable algorithms and models to learn patterns from data and help them to adapt and enhance images based on learned features[9-10]. Here, we explore some of the most popular machine learning image enhancement methods and each contributing to the advancement of visual quality in various applications. Super-Resolution Using Convolutional Neural Networks (CNNs) aims to enhance the spatial resolution of an image that enables sharper and more detailed visuals. CNNs, especially deep architectures like Very Deep Super-Resolution (VDSR), Super-Resolution Convolutional Neural Network (SRCNN) and Enhanced Deep Super-Resolution (EDSR) have demonstrated remarkable capabilities in upscaling low-resolution images. These networks learn intricate patterns and textures from high-resolution training data [11-12]. Generative Adversarial Networks (GANs) is a class of deep learning models that consist of a generator and a discriminator network that work adversarial to improve image quality. In image enhancement, GANs can be

trained to generate images with improved visual quality that often surpassing traditional methods. In particular, Progressive GANs have been successful in generating high-resolution images with realistic details and making them valuable for tasks like image restoration and style transfer [13]. Denoising Autoencoders use unsupervised learning to reduce noise and artifacts in images. These models are trained on noisy images with a clean counterpart and thus, it learns to map noisy input to clean output. With the understanding of patterns in the data, they can effectively remove various types of noise 14]. Image Dehazing Models are type of ML algorithms that has been applied to address the challenge of hazy or foggy images. Dehazing models are usually based on deep neural networks and aim to remove atmospheric haze and improve visibility in images. These models force convolutional layers to capture and enhance features and provide clearer images in conditions where visibility is compromised [15-16].

Some of the other machine learning methods are also discussed here in brief; Image Colorization Networks that colorizes grayscale images is another application of ML in image enhancement. Models like DeOldify and ColorNet utilize deep learning techniques such as GANs and CNNs to predict and add color to black-and-white images. This is particularly valuable for revitalizing historical photographs and improving visual appeal. Low-Light Image Enhancement technique improves the image captured in low-light conditions that is a challenging task. ML models especially those based on deep learning architectures have shown promising results in addressing this problem. These models learn to adaptively enhance details, reduce noise and improve visibility in images with low light [17]. Deep Unfolding Networks combine the power of neural networks with optimization techniques and allows them to iteratively refine and enhance images. These models are often used in tasks like image deblurring and super-resolution where iterative refinement is essential for achieving optimal results. Self-Supervised Learning is the technique where the model generates its own supervision signals from input data. These methods influence pretext tasks such as image rotation prediction or image colorization in order to learn representations that can then be used for various enhancement tasks. Transfer Learning involves training a model on a source task and then fine-tuning it for a target task. In image enhancement, pre-trained models on large datasets can be fine-tuned for specific applications that give a powerful way to control knowledge learned from diverse data sources [18].

Machine learning has significantly advanced image enhancement by providing versatile and effective tools for addressing various challenges in visual data. These methods involve deep learning principles and have demonstrated their expertise in tasks ranging from super-resolution to denoising and dehazing. The adaptability and generalization capabilities of these models make them invaluable across diverse applications including medical imaging, satellite imagery and creative arts. As research in machine learning continues to evolve, the connection between sophisticated algorithms and massive datasets promises even more powerful image enhancement techniques in the future.

IV COMPARISON OF TRADITIONAL AND MACHINE LEARNING METHODS

As the image enhancement is a critical aspect of image processing and is a major aspect to improve the visual quality of images for various applications. Traditional methods that involve mathematical operations and heuristics are considered as the foundation for years. On the other hand, machine learning methods especially deep learning approaches have emerged more recently that includes neural networks to learn complex patterns directly from data. Thus, in order to know more about their approach, performance, data dependency, we have compared both these two approaches across different criterion and the same is displayed with the help of Table 1 shown below.

Table 1. Comparison of traditional and machine learning methods on different basis

Criterion	Method Basis	Traditional Methods	Machine Learning Methods
Approach	Principle	These methods rely on predefined algorithms and mathematical operations.	These methods especially deep learning, learn features and patterns directly from data.
	Process	They typically involve pixel-level transformations, convolution masks or frequency domain operations.	In these methods, the models are trained on large datasets to learn intricate relationships between input and output.
	Flexibility	Limited adaptability to diverse image characteristics.	High adaptability due to the ability to learn complex and non-linear mappings.
Data Dependency	Data Independence	They are designed without reliance on specific datasets.	Highly reliant on the quality and diversity of the training data.
	Generalization	Restricted ability to generalize to diverse datasets and varied image characteristics.	They can generalize well to unseen data if trained on diverse and representative datasets.
Performance	Effectiveness	Effective for simple tasks and scenarios.	They usually outperform traditional methods especially in tasks like feature extraction.
	Complexity	Limited ability to handle complex image conditions such as non-uniform illumination.	These can handle complex scenarios including non-linear mappings and intricate relationships.
Adaptability	Flexibility	Limited flexibility and these methods often require manual tuning for different scenarios.	High flexibility due to the ability to learn from data.
	Scenario-Specific	Well-suited for specific tasks but may not generalize across diverse scenarios.	Capable of generalizing well to various scenarios if trained adequately.
Computational Complexity	Efficiency	They are generally computationally efficient.	They can be computationally intensive especially deep learning models.
	Real time Processing	They are more suitable for real-time applications due to simplicity.	They require significant computational resources which affect real-time applications.
Interpretability	Interpretability	They are more interpretable as operations are explicit and rule-based.	These models can be complex and less interpretable.
	Ability to explain	Easier to explain and understand the impact of each step.	The "black-box" nature of some models sometime limits the understanding.
Robustness	Robustness	These methods can be robust for specific tasks and controlled conditions.	These can be robust across a range of conditions with proper training.
	Noise Handling	Limited ability to handle complex noise patterns.	They can adapt to and handle complex noise patterns.
Parameter Tuning	Parameter Tuning	They often require manual parameter tuning.	Learning algorithms automatically adjust parameters during training.
	User Interaction	They involve trial and error detection for optimal results.	Less user interaction required once trained.
Applications	Applications	Widely used in basic image processing task like contrast adjustment, filtering.	Versatile and thus, applicable to a wide range of tasks including super-resolution, denoising and dehazing.
	Specific Tasks	They can excel in specific, well-defined tasks.	They can adapt to diverse applications with the right architecture and training data.
Evolution and Research	Evolution	Established and well-documented over the years.	Rapidly evolving with ongoing research contributing to frequent advancements.
	Research	Limited ongoing research compared to machine learning.	Active and dynamic field continually pushing the boundaries.

VI CONCLUSION

Finally, the choice between traditional and machine learning methods depends on the specific requirements of the task, the available computational resources and the complexity of the image enhancement problem at hand. This paper provides an in-depth exploration of the conventional as well as advanced techniques of image enhancement covering the concept, implementation needs, advantages and limitations. The discussion spans traditional methods and machine learning-based algorithms that offer a unique perspective on their roles to improve the visual quality of an image. An innovative classification of machine learning-based algorithms based on model strategy is discussed, differentiating them from traditional methods. To facilitate a comparative analysis, various parameters are considered like performance, complexity, efficiency, approach etc. In order to recognize the pivotal role of image enhancement, the approaches are differentiated on different basis to identify the best algorithm for a specific application and which can handle common problems like noise, high contrast and fog. In future, continuing the research in this area aims to force the development of image enhancement algorithms that are not only robust in diverse scenarios but also applicable to a broader range of complex visual tasks in the realm of computer vision.

VII COMPETING INTERESTS

The authors have no Competing interests at stake and there is No Conflict of Interest.

REFERENCES

[1] A Review on Image Enhancement Techniques April 2019, Southeast Europe Journal of Soft Computing 8(1)

[2] Rana, S. B., & Rana, S. B. (2015). A Review ofMedical Image Enhancement Techniques forImage Processing. International Journal of CurrentEngineering and Technology, 5(2), 1282-1286.

[3] Sandbhor, B., & Kharat, G. U. (2015). A reviewon Underwater Image Enhancement Techniques.Int. J. Advanced Research in Computer Scienceand Software Engineering, 5(5), 676-680.

[4] Lepcha, Dawa Chyophel, Ayush Dogra, Bhawna Goyal, Vishal Goyal, Vinay Kukreja, and Durga Prasad Bavirisetti. "A constructive non-local means algorithm for low-dose computed tomography denoising with morphological residual processing." Plos one 18, no. 9 (2023): e0291911.

[5] Jobson, D. J., Rahman, Z. U., & Woodell, G. A.(1996, January). Retinex image processing:Improved fidelity to direct visual observation. InColor and Imaging Conference (Vol. 1996, No. 1,pp. 124-125). Society for Imaging Science andTechnology.

[6] Chong, S. S., Sim, K. S., & Nia, M. E. (2013,October). Modified HL contrast enhancementtechnique for breast MR images. In 2013 IEEEInternational Conference on Signal and ImageProcessing Applications (pp. 360-364). IEEE.

[7] Guo J, Ma J, García-Fernández ÁF, Zhang Y, Liang H. A survey on image enhancement for Low-light images. Heliyon. 2023 Mar 16;9(4):e14558. doi: 10.1016/j.heliyon.2023.e14558. PMID: 37025779; PMCID: PMC10070385.

[8] Litjens, G., Kooi, T., Bejnordi, B. E., Setio, A. A. A., Ciompi, F., Ghafoorian, M., … & Sánchez, C. I. (2017). A survey on deep learning in medical image analysis. Medical image analysis, 42, 60-88.

[9] Jobson, D. J., Rahman, Z. U., & Woodell, G. A. (1997). Properties and performance of a center/surround retinex. IEEE transactions on image processing, 6(3), 451-462.

[10] Wang, Y., Zhou, Q., Liu, J., Xiong, J., Gao, G., & Wu, X. (2019). Proceedings of the 2019 IEEE international conference on image processing (ICIP).

[11] Luo, W., Liu, W., & Gao, S. (2017). 2017 IEEE International Conference on Multimedia and Expo (ICME). [12] Huy, N., & Le Nguyen, M. (2016). Computer vision and pattern Recognition.

[13] Li Chongyi, Guo Jichang, Porikli Fatih, Pang Yanwei. LightenNet: a convolutional neural network for weakly illuminated image enhancement. Pattern Recognit. Lett. 2018;104:1522.

[14] Meng Yingying, Kong Deqiang, Zhu Zhenfeng, Zhao Yao. From night to day: GANs based low quality image enhancement. Neural Process.

Lett. 2019;50(1):799814.

[15] Sobbahi Rayan Al, Tekli Joe. Low-light homomorphic filtering network for integrating image enhancement and classification. Signal Process.

Image Commun. 2022;100.

[16] Goyal, Bhawna, Ayush Dogra, Dawa Chyophel Lepcha, Vishal Goyal, Ahmed Alkhayyatd, Jasgurpreet Singh Chohan, and Vinay Kukreja. "Recent advances in image dehazing: Formal analysis to automated approaches." Information Fusion (2023): 102151.

[17] Gonzalez RC, Woods RE. 2002. Digital image processing.2nd ed. Englewood Cliffs, NJ: Prentice-Hall. ISBN: 0-201-18075-8.

[18] Raju G., and Madhu S. Nair. 2014. "A fast and efficientcolor image enhancement method based on fuzzy-logic andhistogram." AEU-International Journal of Electronics andCommunications 68.3 : 237-243.