Enhanced Morphological Model for Enhancement of Dermoscopy Images for Edge Based Segmentation

Tumors form when human skin cells divide and grow unevenly; melanoma, squamous cell carcinoma, and basal cell carcinoma are the three main types of skin cancer [1]. According to the data, the predicted incidence of skin cancer is 35 percent. The fact that half of the 20 million cancer cases reported globally in 2023 were fatal lends credence to the severity of this disease. Melanoma, a form of skin cancer that starts in melanocytes, grows quickly and is very dangerous [2]. Melanocytes are specialized cells whose primary role is to produce melanin. The proliferative potential of the cell is drastically reduced during this differentiation operation.

Examining a skin lesion traditionally has involved measuring its size, shape, and formation manually [3]. Because professionals and dermatologists need to physically perform these operations, they are less precise and take more time. It is crucial to prioritize the early detection of skin cancer in order to control the patient fatality rate [4]. On the other hand, these differences are being gradually addressed by Artificial Intelligence (AI) models. In recent years, new deep and machine learning approaches have been established that are based on computer vision. These cutting-edge methods allow medical professionals to diagnose and categorize these ailments with the help of a machine [5]. As a result, getting a computer-aided prognosis is crucial for getting better results and more accurate results.

Each skin lesion must first be preprocessed and its boundaries estimated before features can be extracted and used for lesion classification [6]. For the last ten years, researchers have been preparing medical images using a variety of methods. Artefact removal, color normalisation, and contrast stretching are all examples of preprocessing approaches [7]. While these preprocessing models do their jobs well, they add significant processing time to the algorithm, which impacts both the training and testing phases [8]. In order to improve the classifier's performance, the dermoscopy images are preprocessed with the boundary estimation procedure, which separates the borders of the images. In order to recover the Region of Interest (RoI) from surrounding background lesions, the segmentation model must exist. Furthermore, this method aids in the improved recognition [9] of the intrinsic clinical aspects of skin lesions. The accuracy of melanoma prognoses is correlated with the precision of segmentation algorithms. In order to improve the classifier's overall accuracy, segmentation approaches are typically the backbone of most research models [10].

Due to the many databases of skin lesion types, each with its own unique complexity, the procedure remains an open task, leaving lots of room for future research, even after complex segmentation models have been proposed [11]. False classifications and accuracy issues caused by these asymmetric features make segmentation models more difficult to implement [12]. Changing the brightness, contrast, levels of distortion, and lightness in the dermoscopy images also drastically reduces the segmentation accuracy [13]. More effective ways to deal with these practical challenges are thus greatly needed. The aberrant reinforcement of certain cells that occurs as a result of changes in gene expression on the skin layer is the defining feature of skin cancer. These malignant cells spread to neighboring cells. Because the human body is susceptible to a wide range of influences in the modern world, including, but not limited to, increased life expectancy and exposure to ultraviolet radiation, the number of cancer patients has risen dramatically [14].

Malignant and benign skin cancers are the two main categories. Metastasis, the process by which cancer cells travel to other parts of the body, distinguishes the two types. Cancers that are considered malignant have the ability to penetrate and kill off nearby tissues. Additionally, it can travel through the lymphatic system and bloodstream to distant organs and tissues [15]. A large percentage of the population is affected by this disease. By putting pressure on nearby nerves or blood arteries, even benign cancer, which is more limited, can impact the environment. Benign cancers tend to grow at a slower rate than malignant ones [16]. Negligible treatment of these malignancies may have devastating consequences for human health. Preliminary testing for skin cancer is, hence, essential for an accurate diagnosis. The biopsy is an intrusive procedure that causes pain and discomfort for cancer patients [17]. The goal of dermoscopy imaging is to examine the skin layers in great detail using a microscope and other specialized lighting equipment in order to prevent the need for an unneeded biopsy. The most challenging aspect of dermoscopy is identifying skin lesion types and confirming their existence in pictures. Segmentation, feature extraction, and the classification process are the few stages needed for skin lesion detection [18]. This research proposes an automatic segmentation method that can be used as a first step in skin lesion classification. Dermatologists find this automated approach useful for detecting skin cancer since it identifies and locates the areas of skin lesion.

1.png

Figure 1. Dermoscopy images

Researchers have paid a lot of attention to computer-aided technologies that can analyze medical images for diagnostic purposes. These are specifically created and adjusted to help with things like segmenting and classifying the ROI, which in this case includes areas with cancer. As a general rule, cancer typically has a delayed clinical beginning [19], therefore early detection and delimitation of lesion boundaries are crucial for effective treatment of the disease, especially in its early stages. Nearly 17 million individuals are impacted by cancer each year, with approximately 9.6 million losing their lives as a result of treatment delays. As a result, cancer is now the top killer on a global scale. One of the most common malignancies in both children and adults, skin cancer develops in the skin's outer layer called epidermis [20]. For the purpose of detecting cancer boundaries from dermoscopy images, several computer-assisted methods have been suggested. The dermoscopy images is shown in Figure 1.

In addition to being the most common form of skin cancer, melanoma is also the most aggressive and lethal because of its high metastatic rate. A malignant skin cancer known as melanoma occurs when melanocytes [21], the skin's pigmented cells, grow irregularly. Cancer can start anywhere on the skin's surface, spread to other parts of the body, and even start in the chest or back. This skin cancer has the highest fatality rate compared to all others, and its incidence rate has been steadily rising, reaching 4-6% every year. The five-year survival rate can reach as high as 98% with an early diagnosis. Given the statistics surrounding melanoma incidence and mortality rate, it is crucial to diagnose patients promptly in order to provide them with effective therapy [22].

Among the many possible approaches to the issue of digital image enhancement, mathematical morphology stands out. For every pixel in the processed image, these operators choose a new grayscale value between two patterns based on a proximity metric [23]. Alternatively, the homomorphism filter operates in the frequency domain, and nonlinear functions like logarithm or power functions are among the most used approaches in image processing for improving dark areas [24]. One major drawback of histogram equalization is that it often fails to preserve details well because it applies the image's global attributes incorrectly in a local context. The normal and melanoma images are shown in Figure 2.

2.png

Figure 2. Normal and melanoma images

3.png

Figure 3. a) Normal image; b) Segmented Image; c) Extracted portion

Segmenting images is an essential step in most computer vision, video, and image applications. Partitioning an image into sections that should ideally represent distinct items in the actual world is a common usage for it. Content analysis and image comprehension rely on it. It is difficult to compare several segmentation methods or even alternative parameterizations of a single approach since there is still no adequate performance measure, despite the development of numerous segmentation methods [25]. Segmenting an image into its component components and then extracting the items of interest is a common way to explain image segmentation. The dermoscopy image segmentation process is shown in Figure 3.

The outcomes of segmentation have a profound impact on all the subsequent steps of image analysis, including object representation and description, feature measurement, and even higher-level tasks like object classification and scene interpretation, making it a crucial component of automatic image analysis [26]. The process essentially entails dividing a digital image into various zones that correspond to specific surfaces, objects, or intrinsic object features. It entails classifying each pixel, then finding clusters of pixels that have certain visual traits or areas that are similar to one another. Each of these areas may, ideally, represent an object or pattern in the image. This research proposes a Multi-Level Image Quality Enhancement Model using Enhanced Morphology Model with Edge-Based Segmentation (MLIQE-EMM-ES) for accurate detection of melanoma in dermoscopy images.

One goal of automated visual assessment in medical imaging is the early detection of certain disorders. The key to successful treatment and, more importantly, a full recovery from potentially fatal diseases is early detection. In the absence of prompt medical attention, many disorders have the potential to progress and even kill. Because skin lesions are initially too small to be properly identified by human vision, automated visual inspections with computer-aided diagnostic systems supplement a human expert examination, which is crucial for the early detection, diagnosis, and treatment of abnormal cell development in humans’ skin [27]. Cancers of the skin that arise from aberrant skin cell development fall into three main categories: melanoma, basal cell carcinoma, and squamous cell carcinoma. Although it is uncommon, melanoma is the leading cause of death among skin cancer patients. The consequences can be catastrophic if it spreads to deeper tissues and organs. Dermoscopy is a technique for skin lesion segmentation, pattern detection, and classification that uses magnification and the elimination of skin reflection [28].

4.png

Figure 4. Dermoscopy image segmentation

Because they look so similar, skin lesions can be difficult to distinguish between benign and malignant. Imperfect borders, artificial markings, hair artifacts, and background fluctuations are only a few of the drawbacks [29]. Basic thresholding methods with simple objective functions can be utilized for segmentation when working with images that have clear borders and no artifacts. Finding the optimal thresholding value, however, requires a well-defined objective function when images display a great deal of fluctuation [30]. It is worth noting that prior studies have mostly concentrated on utilizing various machine learning approaches, ignoring the usage of thresholding methods. The segmented dermoscopy image is shown in Figure 4.

When diagnosing a condition and tracking the efficacy of a treatment plan, medical imaging plays a crucial role. The resulting images could not be of sufficient quality for a precise diagnosis, despite the fact that methods of obtaining these images are becoming more advanced. Images may have inferior quality for a variety of reasons, including but not limited to: technical limitations of imaging instruments, patient unique conditions in images, surrounding sounds, and emergency scenarios. When reimaging is not an option, image enhancement techniques can be a lifesaver. The damaged images are repaired and their quality and contrast are enhanced using these new techniques. One method of image processing that relies on an object's shape is morphology. In order to generate a new image of the same size from an input image, morphological approaches apply a structuring element. A comparison of the matching pixel in the input image with its neighbors is used to determine the value of each pixel in the input image, create a morphological procedure that can identify and respond to particular shapes in the input image by adjusting the neighbor's size and shape. It is possible to establish the morphological procedures on grayscale images with a planar single-channel source image. Afterwards, the scope can be broadened to encompass full-color pictures. The morphology operations on an image are shown in Figure 5.

5.png

Figure 5. Morphology operations

6.png

Figure 6. Proposed model framework

One crucial stage in the computer-assisted diagnosis of melanoma is the automatic segmentation of skin lesions on dermoscopy images. However, there are large individual differences in the appearance of lesions, making this a difficult process. When working with massive amounts of image data, this problem becomes much more complex. To further increase the segmentation performance, color information is explicitly used from several color spaces. This helped with network training. When doing image analysis, segmenting the image is the initial step. The process of segmentation splits an image into its individual elements. The problem at hand dictates the extent to which this subdivision is pursued. Two segmentation techniques the discontinuity detection approach and the similarity detection technique are available in case the item needs to be separated from the backdrop in order to read the image properly and determine its content. Partitioning an image according to sudden changes in gray-level image is one way to go about the first method. In the second method, the expanding threshold and region serve as the foundation. The proposed model framework is shown in Figure 6.

An additional method for precise melanoma diagnosis in dermoscopy pictures is the combination of erosion and dilatation with edge-based segmentation. As effective preprocessing procedures, morphological erosion and dilation can eliminate minor bright artifacts like residual hair, specular reflections, and noise, while dilation can fill in tiny gaps and rejoin broken lesion edges. These procedures, when applied in conjunction with one another during opening and closure procedures, eliminate rough edges around lesions, hide imperfections in the background, and protect the lesion's general structure. Performing this preprocessing step guarantees a clearly defined lesion region prior to implementing more accurate border detection techniques. Next, the cleaned image is subjected to edge-based segmentation, which can identify distinct changes in intensity between the lesion and the surrounding skin. With much of the unnecessary high-frequency noise removed by morphological filtering, edge detection yields more precise and clearer boundary maps. The combination of morphology's lesion isolation and edge detection's fine shape refinement yields pixel-level segmentation precision, creating a synergistic effect.

The flexibility and ability to integrate with other processing stages are two key ways in which enhanced morphological procedures differ from classical ones. Conventional morphology ignores local image features in favor of applying global, fixed-structure elements such as opening, closure, dilation, or erosion. A contrast to this is the MLIQE-EMM-ES model's enhanced morphological approach, which uses multi-scale structuring elements with different radii to capture coarse and fine lesion features. To make sure the improvement is concentrated where it is most required, it uses pixel-intensity analysis to preferentially target low-contrast pixels rather than applying procedures randomly. In addition, improved morphology uses local pixel normalization to dynamically modify intensity ranges, which improves contrast in poorly defined lesion edges. The technique improves lesion-to-background separation by integrating white top-hat procedures, which enhance bright features, with bottom-hat operations, which emphasize dark features. Especially for difficult low-contrast melanoma images, this method maintains lesion information, suppresses noise, and generates high-quality segmentation outputs when integrated into a multi-level processing pipeline including edge detection.

The Pseudocode for the proposed model is clearly discussed.

First, the model-based design employed for image pre-processing is an aspect of the edge detection technique. For subsequent processing, the 2D data needs to be translated to 1D, and a color space conversion block does just that, turning RGB into a grayscale image. The second method is morphological operations segmentation, which is different from the first. Objects' shapes and forms are examined in morphological processes. Image segmentation, which involves separating related items, object extraction which involves removing small objects or noise from an image, and measurement operations which include texture analysis and shape description are all under the purview of morphological image analysis. Elements of the Image Processing Blockset can be used to execute morphological operations like opening, closing, dilation, and erosion. For morphological image analysis, a mix of these blocks need to be considered. Users can filter images, segment them, and quantify them all with morphological image analysis. Using dermoscopy image segmentation, skin degradation can be more accurately recognized. Reliable skin lesion detection is difficult for a number of reasons, including as the image capture method, the characteristics of the lesion, and the skin's texture. Currently, most segmentation datasets rely on noisy expert annotations to represent skin lesion borders because accurate annotations are expensive and time-consuming to create. For dermoscopy-based lesion localization and accurate skin lesion type identification, segmenting the lesion border is crucial. This research proposes a MLIQE-EMM-ES for accurate detection of melanoma in dermoscopy images.

Algorithm MLIQE-EMM-ES

{

Input: Dermoscopy Image Dataset {DIset}

Output: Features Extracted Set {FEset}

Step-1: Initially, the images from the dermoscopy dataset is considered and these images will be processed by considering all the pixels. The pixels in the image will be analyzed and its pixel contrast levels are considered for image processing. The image pixel contrast processing is performed as

$\begin{aligned} \tau[\mathrm{M}]=\sum_{\text {img}=1}^{\mathrm{M}} & \frac{\sum \text {imgattr}(\text {img})}{\mathrm{M}}+\text {getminInten}(\text {img})+ \text { getmaxInten(img)}+\operatorname{diff}(\text {img}(\mathrm{x}, \mathrm{y}), \text {img}(\mathrm{x}+1, \mathrm{y}+1))\end{aligned}$

$\begin{gathered}\operatorname{IPcontr}[\mathrm{M}]=\sum_{\mathrm{img}=1}^{\mathrm{M}} \text { getImgattr }(\mathrm{img})+\tau(\mathrm{img})+\operatorname{minInten}(\mathrm{img})\end{gathered}$

where, τ is the model that considers the pixel contrast, minInten() model considers the pixels with the minimum intensity levels and maxInten() model considers the pixels with maximum intensity contrast levels. getImgattr() is used to extract the image attributes like color, texture, shape.

Step-2: The process of dilation enlarges objects and fills up minute imperfections. Erosion removes insignificant particles, leaving behind only the actual objects. The two most fundamental morphological operators are dilation and erosion. Dilation chooses the area around the structuring element with the brightest value, whereas erosion chooses the area with the darkest value. The morphological operations are applied on the image as

Step-3: After applying the morphological operations on the images, each image will be analyzed with pixel contrast dissimilarities and such pixels are only considered so that enhanced morphology operations will be applied on them. The enhanced morphology operations are applied as

$\begin{aligned} \text {Emorp}[\mathrm{M}]=\prod_{\text {img}=1}^{\mathrm{M}} & \text { getattr(Erosion(img)) }+\operatorname{getattr}(\text {Dilation(img)})+\gamma(\text { Erosion(img) })+\beta(\text { Dilation(img)}) \\ & +\max (\operatorname{diff}(\mathrm{x}, \mathrm{x}+1))+\max (\operatorname{diff}(\mathrm{y}, \mathrm{y}+1))+\gamma(\text {Opening}(\text {img}))+\beta(\text {Closing}(\text {img}))\end{aligned}$

Here γ is the model that considers the pixel with poor contrast among the processed image and β is the pixels with high contrast so that pixel normalization is performed that improves the image quality.

Step-4: The image quality will be enhanced using the enhanced morphology operations. The edge detection will be performed on the images where the exact shape of the skin lesion will be detected. The exact edges of the image will be helpful for easy and accurate detection of melanoma in the images. The edge detection is performed as

$\begin{aligned} & \text {EdgeSet}[\mathrm{M}]=\frac{\operatorname{getminrange}(\operatorname{Img}(\mathrm{i}))+\beta(\mathrm{img})}{\max \left(\sum_{\mathrm{i}} \operatorname{Img}-\operatorname{Emorp}(\operatorname{Img}(\mathrm{i}))\right)} +\sum_{i=1}^N\left\{\frac{\mid \max \operatorname{range}(\gamma(\operatorname{Img}+\operatorname{Emorp}(\operatorname{img}(\mathrm{x}, \mathrm{y}))| |}{\mathrm{M}}\right\}^{\mathrm{x} * \mathrm{y}}\end{aligned}$

Step-5: Segmentation is applied on the images after edge detection. Segmentation divides the image into multiple portions as only relevant portion is extracted. The features from the relevant regions will be considered for feature processing for melanoma detection. The segmentation is performed as

$\begin{aligned} & \text {Iseg}[\mathrm{M}]=\sum_{i m g=1}^{\mathrm{M}} \sqrt{\frac{\sum_{\text {img}=1} \operatorname{getattr}(\text {EdgeSet}(\text {img}))+\operatorname{maxPix}(\text {img})+\operatorname{minPix}(\text {img})}{(x * y)}}+\prod_{i m g=1}^{\mathrm{M}} \max (\operatorname{simm}(\operatorname{EdgeSet}(\mathrm{x}, \mathrm{x}+1)))+\max (\operatorname{simm}(\operatorname{EdgeSet}(\mathrm{y}, \mathrm{y}+1)))\end{aligned}$

Here simm() model is used to check the similarity levels of the edge feature patterns.

}