Patibandla Sandhya Krisha*
| Subba Rao Peram
© 2023 IIETA. This article is published by IIETA and is licensed under the CC BY 4.0 license (http://creativecommons.org/licenses/by/4.0/).
OPEN ACCESS
Oral cancer, the most prevalent form of head and neck cancer, calls for early detection to ensure better patient outcomes, reducing morbidity and mortality rates. This study explores the application of computer vision and deep learning methods for photographic images in the oral cancer domain, investigating a two-stage pipeline for an automated system to identify oral potentially malignant abnormalities. Oral cancer staging, crucial for determining appropriate treatment and medication, often faces challenges due to noise levels in images that impact disease prediction accuracy. This research works with an image dataset, enhancing image quality and performing denoising to improve accuracy levels. The study aims to evaluate the accuracy of an image enhancement and denoising model, resulting in quality images for extracting features for oral cancer detection. By segmenting the image using multiscale morphology methods, cell features can be extracted. The morphological edge detection method enables more precise extraction of target, cell area, perimeter, and other multi-dimensional features, followed by classification through Convolution Neural Networks (CNN). This research proposes a Precise Denoising Model with Edge-Based Segmentation for Labeled Pixel Extraction with Fixed Feature Set (PDM-ES-LPE-FFS) for relevant feature extraction. When compared with traditional models, the proposed model demonstrates superior performance.
oral cancer, image processing, segmentation, denoising, edge detection, pixel extraction, image quality enhancement, feature extraction
Oral cancer is on the rise around the world as a result of the modernization and civilization. In the most recent global census, mouth and throat cancers ranked sixth among all neoplastic conditions [1]. Oral cancer seems to have a 5-year survival rate with only 35–40%. Professionals have made significant advancements in the treatment of oral malignant tumours over the years. Even so, there has been no major increase in the survival rate of these patients in the last few decades. In part, this is due to a lack of public awareness of oral cancer [2], which results in patients not paying enough attention to early signs and delaying the best possible treatment, resulting in irreversible damage [3]. It is therefore critical that oral cancer be diagnosed as early as possible in order to improve the cancer's cure rate and eradicate the tumour [4].
Oral Potentially Malignant Disorders (OPMDs) are oral lesions that can be found during normal screening by a clinical oral examination carried out by a general dentist, therefore a delayed diagnosis is not necessarily a deal breaker [5]. Once an abnormality is found, the patient will be sent to an expert for additional evaluation. Early detection, down-staging of the disease, and a reduction in mortality have been seen in previous trials in India, where screening has been used. Because of the scarcity of oral cancer specialists and treatment options in low and middle income countries, screening programs must provide a low-cost and quick way to detect the disease [6]. Telemedicine could be a potential option in this situation. Experts using a clinical oral examination and those using photos taken with cell phones had a moderate to high degree of agreement in their clinical diagnoses [7].
Figure 1. Oral CT images
Specialists may be able to improve screening programmes' recommendation accuracy by providing remote consultations [8]. Incorporating an automatic detection method linked to machine learning to evaluate mobile phone images is a terrific idea. Microscopic pictures play a major role in the automatic detection of oral cancer, OPMDs [9], and benign lesions. Multidimensional hyperspectral images, CT (computed tomography) [10], autofluorescence, and fluorescence imaging, as well as normal white light images, can be considered for the tumor detection. The CT images of oral region is show in Figure 1.
Digital image processing is the process of applying various techniques and handling to photographs in order to achieve a specific aim. There are many different components to digital image processing, such as picture digitization [11], application uses, image processing, image output, and display. Image digitization is the process of converting an image into a set of numeric values, which may subsequently be analyzed and interpreted by a computer or other digital devices [12]. The process of transforming an image into a digital format is known as digitization. An integer number is generated by sampling and quantizing the image's brightness at every pixel point [13]. Once all pixels have been converted, the image is stored in the computer as an integer matrix. Pixels in a photograph have two properties: their position and their colour [14]. Grayscale integers, commonly known as row and column coordinates, are used to describe the pixel's location in relation to other pixels on the scan line [15]. The most often employed digital instruments for image processing are flying spot scans and microdensity metres [16].
Digital pathology involves the digitization of tissue images and slides. Tissue digitization could improve the efficiency of routine diagnostic pathology workflow by making it easier to store, view, and examine tissue pathologically [17]. Cancer cell digital pathology relies heavily on preprocessing, image segmentation, extraction of features, and classification methods. Pathologists can determine the cancer's malignancy or benignity by examining microscopic photographs of the cancer cells [18]. Useful feature parameters can be extracted and identification rates improved with cell wavelet transform. It is common to use mathematical morphology in the analysis of microscopic cell images [19]. Mathematical morphology can help with digital picture analysis [20]. Obtaining morphological information about a target is as simple as determining its size, shape, orientation, and connectivity. The oral tumor samples in the general images are shown in Figure 2.
Figure 2. Oral tumor samples
CT scans serve as the basis for radiation therapy treatments. Most doctors prefer CT imaging, and it may also be used to measure human biological indicators, such as blood pressure and heart rate. Oropharyngeal carcinoma can be detected using CT scan images [21]. Image processing is part of the process of creating a photograph. Due to the inherent noise and irregularity of CT scans, filtering techniques such as Anisotropic Diffussion Filter, Gaussian Filter, and Adaptive Median Filter can be employed to reduce speckle noise and improve image quality [22]. Measures such as PSNR and MSE are used to assess how well these strategies perform [23].
Image denoising, the process of reducing noise from a noisy image, can be used to restore an original image. Images that have undergone denoising may lose some information since it is challenging to distinguish high frequency noise, edges, and textures [24]. High-quality image capture is a critical need in the modern world, and one solution is to eliminate noise from otherwise chaotic shots. After many years of study, image denoising is now a well-known problem. Nonetheless, we still face formidable obstacles. As image denoising is an inverse math problem, there is no one correct answer.
Technical limitations of the camera's image sensor or poor environmental conditions are common causes of noise, which might include random variations in brightness or colour information. Image noise is a widespread problem that needs to be dealt with the right denoising techniques in real-world situations. Denoising an image is a complex process since the high-frequency content of the image, i.e., the details, is related to the noise. It is therefore important to establish a compromise between minimizing noise as much as feasible while retaining as much information as possible. Inverse, Median, and Wiener Filters [25] are the most used methods for removing noise from images.
Splitting a picture into groups based on comparable qualities and attributes is known as image segmentation. Segmentation of an image is defined mathematically as a finite set of regions. Research categorizes segmentation methods based on a variety of features, such as the region, entropy, shape, threshold, and pixels' association, among others. CT images were used to identify specific locations or areas of interest. Dentists use a variety of approaches to segment oral images, including region, cluster, threshold, border, and watershed segmentation [26].
Using deep learning and linked component analysis, researchers have been able to discover, classify, and measure patterns in radiographs. The ability to analyze hierarchical feature representations gained from data is at the heart of these advancements. Many image analysis pipelines rely on the analysis of connected components [27]. Foreground and background pixels are separated in an image. Finally, foreground and background pixels are linked together and removed, while border pixels are linked together to produce image borders. After the picture has been preprocessed and threshold, image regions analysis is used.
There are numerous applications for CNN that go beyond generic denoising, including blind denoising, real-world noisy images, and many more. Academics are still working on a complete evaluation of CNN algorithms for denoising images [28]. Denoising techniques for images have been classified by CNN based on the sort of noise they encounter. This research proposes a Precise Denoising Model with Edge based Segmentation for Labeled Pixel Extraction model for relevant feature extraction.
As in Magnetic Resonance Imaging (MRI), Compressive Sensing (CS) offers a platform for the acquisition of slow and sequential data. In order to use it for high-speed data consideration, CS has to be very computationally intensive in order to rebuild MR pictures from sparse k-space data. Removal of Impulse response noise from magnitude MR images, which alters the image properties and hence impacts the clinical utility, is another key hurdle that needs to be overcome. Gaussian noise models have been used in most of the research so far. Rician noise models, in particular, are rarely used in the CS paradigm. An entirely new framework for reconstructing MR images from noisy sparse k-space data is developed in this work. A convolutional neural network (CNN) was proposed by Manimala et al. [2] to denoise MR pictures distorted by Rician noise. Using signal similarities, the algorithm processes similar patches as a group in order to extract local feature sets. With Convolutional Infrastructure for Fast Feature Embedding framework, the CNN was trained on a GPU for online reconstruction, which resulted in a significant reduction in run time. Because it does not require optimization or prediction of sound levels while denoising, the CNN-based reconstruction has a significant advantage over currently available state-of-the-art approaches.
The concept of picture denoising has taken a particular interest in deep convolutional neural networks (CNNs). However, there are two drawbacks: training a deeper CNN for denoising tasks is extremely difficult, and the majority of deeper CNNs suffer from efficiency saturation. An innovative network, the batch-renormalization denoising network, was developed by Tian et al. [3]. The author specifically joined two networks in order to enhance the network's width and so gain more features. Batch renormalization is integrated into BRDNet so that the internal covariate shift and small mini-batch issues can be dealt with. The network training is made easier by including residual learning into the process on a broad scale. Using dilated convolutions for denoising, more information can be extracted.
Manimala et al. [4] introduced a new approach for MR image reconstruction called Adaptive Sparse Reconstruction utilizing Convolution Neural Network (AsrCNN) in order to improve picture quality and speed up reconstruction. Four convolution layer and one fully linked layer make up the AsrCNN. The proposed approach is trained on a large dataset with adaptive gradient optimization, resulting in high-quality reconstructions of MR images. For the dictionary, the training set includes 3×2, 3×2, 3×2 images with 32 weights, which are utilised to update the weights. It is used to recover sparse MR images that are distorted by Gaussian noise. AsrCNN's patch-based technique eliminates the need to resize MR pictures of various sizes.
There is a constant search for an effective image denoising technique at the intersection of statistical and functional analysis. According to Goyal et al. [9], there are numerous denoising techniques that can be used in grayscale imaging, however the level of usefulness of these algorithms still has room for improvement. There are several instances where Gaussian noise deters information from the image pixels. All approaches operate best when predicated on a certain set of assumptions, but in general, they tend to introduce artefacts and obliterate delicate structural features.
Deep learning (DL) techniques have been used in a variety of medical imaging applications, including classification, segmentation, and detection. These approaches are used for denoising, enhancing, and restoring images. DL-based methods Image denoising is a critical step before performing any image analysis. Image denoising has recently seen a rise in the use of deep learning techniques to produce the best possible results. Autoencoder models for medical image denoising were used by Nasrin et al. [10] for digital pathology, dermoscopy, Magnetic Resonance Imaging (MRI), and Computed Tomography (CT) image denoising in this research. TD transfer between MRI and CT scan images is also tested to see how well the R2U-Net autoencoder model performs.
Denoising approaches for biological microscope pictures are reviewed and a novel and original sparsity-based algorithm is introduced in this study. According to Meiniel et al. [17], a single set of parameter values and no prior model can be used to manage both basic and complicated types of noise without the need for an a priori model or a large number of parameters. Denoising performance of thirteen cutting-edge approaches, each tailored to deal with a distinct sort of noise in bioimaging, is also evaluated in an extended comparison.
Speckle reduction is still a major concern in ultrasound picture processing and analysis. Due to its superior despeckling performance, the nonlocal means (NLM) filter has recently attracted a lot of interest. In order to use the gray-level information from a noisy image to assess the similarity between two patches, most known NLM algorithms currently use. This makes it challenging to accurately depict structural similarity in ultrasound images. Yu et al. [19] presented an NLM technique based on PCANet, a simple deep learning baseline. Rather than using pixel intensities to compute pixel similarity, Mafi et al. [21] used the intrinsic properties of picture patches generated by this network. PCANet's original binary hashing and block histograms are replaced with a parametric Rectified Linear Unit (PReLU) activation function in this technique. The convolution kernels of this model are first learned by training it on the ultrasonic database. This pre-denoised version of the noisy image will be used to train the PCANet, which will then extract intrinsic features from the image patches. Using these features, the structural similarity between picture patches is determined in the NLM approach and the weighted average of all pixels in a search box is computed to form the finished despeckled image.
Deep convolutional neural networks can be utilised for picture denoising because of their versatile designs. The following limitations, however, make them unsuitable for this use: It is extremely difficult to train a deep neural network. In the case of larger networks, the problem of performance saturation becomes more acute. Enhanced convolutional neural denoising network (ENDN) was proposed in this study by Tian et al. [25]. Remaining learning and batch normalisation approaches are used to deal with training issues and speed up network convergence. Additionally, the suggested network makes use of dilated convolutions to boost context information while also lowering computational costs.
Images are vulnerable because of the various sorts of noise that can distort the graphic information recorded inside them. Image de-noising has been a part of digital image processing from the beginning. Its goal is to eliminate noise in the background while bringing attention to the visual information it carries. For denoising, deep learning is an essential tool because of its durability, accuracy and speed. Many contemporary state-of-the-art image denoisers, such as dictionary learning models and convolutional neural networks are examined for a range of noises, such as Gaussian, Impulse Poisson and Mixed.
Analysis models reveal the forwarded de-noising model to the user, and the solution method is selected based on predetermined criteria. Depending on the image nature, spatial filter deterministic model has various difficulties. Edge erosion and blurring are common artefacts of spatial and transform domain techniques. For the development of an inverse model that can be applied to a new dataset, deep learning models require picture datasets that comprise both clean and noisy images. Analytical approaches to data de-noising rely on an optimization process that is computationally costly and on the heuristic selection of hyper-parameters to reduce errors. As a result of the feature learning process, deep learning models have been proven to exceed analytical methods in terms of performance.
Denoising filters rely heavily on the identification of noisy pixels in order to increase their performance. Noisy pixel detection is absolutely necessary for edge protection. Filter efficiency is reduced when there is a mismatch between noise-free and noisy pixels. Original pixel values are affected by the level of noise, which results in a minimum or maximum grey value. The original pixels of some photographs may have a grey value between 0 and 255. Denoising filters may reduce the effectiveness of the original pixels if they are mistaken for noisy pixels. This three-value filter uses the minimum, medium, and maximum values of pixels in the processing window to characterise the pixel class and recreate noisy pixels. This filter enhances a noisy image by adding a window of size 5 * 5. From top to bottom and right to left, the window moves. The similarity between a noisy pixel and a noise-free pixel is analyzed.
Images are segmented into zones, and the object of interest is extracted from each segment. This is a crucial stage in image processing and subsequent picture interpretation. The most fundamental and important component of computer vision research is computing at this level. Analysis, recognition and interpretation are directly influenced by quality picture segmentation results. An effective image segmentation method can be used to extract features and attributes from the target image, allowing for more advanced image analysis. This revelation is critical since image processing mainly relies on segmentation. The morphological characteristics of the nucleus mostly decide whether a cell is cancerous or not. This is the first and most critical stage in identifying cancer cells, and it is the most important phase. Images are segmented according to a set of predefined parameters. Reasonable segmentation criteria must be applied to all regions and pixels that have relevant properties of pixels in each picture region. The normal CT image and the denoised image are shown in Figure 3.
Figure 3. Noisy and original image
Restoration, visual tracking, registration, segmentation, and classification are just few of the many uses for image denoising, all of which depend on being able to recover the original image content for optimal performance. Noise in the capturing of a MRI image is a common problem. Denoising is a technique used to clean up MRI images by eliminating unwanted background noise while preserving crucial parts of the image content. When images are captured by a machine, the data is typically noisy. It is necessary to perform denoising to extract the relevant features for detection of oral cancer accurately.
The process of extracting visual features from a MRI image is known as feature extraction. The purpose of feature extraction is to provide a simplified representation of the original oral MRI image for use in subsequent decision-making processes, such as classification tasks. Each of the five key aspects of image quality includes contrast sensitivity, detail, noise, artefacts, and spatial set that can be fine-tuned by adjusting the values of the corresponding protocol elements. MRI is a useful diagnostic tool because it may be adjusted to clearly show a variety of medical disorders.
The CT image integrity and connectivity are well preserved after lowering the image noise. The nucleus of a tiny cell and the backdrop have drastically distinct shades of grey at this point. In order to accurately estimate a image border, an edge based segmentation method is used. This study applies the most frequent method of threshold segmentation, which is the maximum variation between classes. The morphological edge detection method, followed by classification using Convolution Neural Networks, allows for more exact extraction of target, cell area, perimeter, and other three-dimensional properties (CNN). In this study, a Precise Denoising Model with Edge based Segmentation for Labeled Pixel Extraction with Fixed Feature Set (PDM-ES-LPE-FFS) model for relevant feature extraction is proposed.
Algorithm PDM-ES-LPE-FFS
{
Input: Oral cancer CT Image Dataset {OCIDS}
Output: Denoised Image {DIS}
Step-1: Initially the oral cancer image dataset is considered for processing and the images are processed for noise removal. The denoised images are considered for the noise removal and the images with high quality are directly processed for feature extraction. The noisy images are considered from the dataset as
$\begin{gathered}D N I D S[M]=\sum_{I=1}^M \operatorname{getImg}(I)+ \text { getIntensity }(I)+T h\end{gathered}$ (1)
Here Th is the threshold value of intensity considered to add to the original image intensity for quality enhancement.
Step-2: The image segmentation is performed to divide the image into multiple portions. The process of image segmentation is widely used in digital image processing and analysis, and it often divides a picture into various areas or segments according to the characteristics of the image's pixels. The image segmentation is based on edge based segmentation technique and the object pixels in each segment is extracted as
$\begin{align} & ISeg(I(X,Y))= & \sqrt{\frac{\sum\nolimits_{I=1}^{R}{(getIntensity(X,Y)+Th)+{{({{(X*Y)}_{M}})}^{{}}}}}{(X+Y)+\theta (X)}} \\\end{align}$ (2)
$\begin{align} & \delta =& \frac{\sum\nolimits_{I=1}^{R}{\max Intensity(I)-\min Intensity(I)-Th+getIsegrange(I)}}{size(Iseg(I)} \\\end{align}$ (3)
$PixSet[M]=\sqrt{\frac{\sum\nolimits_{I=1}^{R}{\delta {{(X)}^{2}}+G(X,Y)-ISeg(I(i))}}{\max (\delta )}}$ (4)
The R is the total pixels considered in a segment, angle of the segment is indicated by ⍬ and X,Y are adjacent extracted pixel set. G is the function used to consider the corrupted pixel set in an image.
Step-3: The extracted pixels in each segment is considered and the standard mean of pixels are considered for identification of dissimilar pixels that are used for oral cancer detection. The standard mean pixel set is calculated as
$\begin{align} & StdPixSer[M]=& abs\,\left( \frac{\sum\limits_{X=1\,}^{R}{\sum\limits_{Y=1}^{R}{\operatorname{Im}gseg(X)-\delta +\max Intensity(FpixSet(I))}}}{\left( X*Y \right)+\theta } \right) \\\end{align}$ (6)
Step-4: The standard pixel set identifies the dissimilarities in the segment and the labeling is performed on the similar and dissimilar pixel set. The labeling of standard pixel set is performed as
$LabPixSet[M]=\left\{ \begin{matrix} 1 \\ 0 \\\end{matrix} \right.\,\,\begin{matrix} (X,Y)\le \theta (\max (\delta )){{\vee }_{X,Y}}\ge X+Y \\ other\,wise \\\end{matrix}$ (7)
Step-5: The labeled pixel set is considered for noise removal so that the quality image features can accurately predict the oral cancer. The denoising is performed as
$\begin{align} & Denoised\operatorname{Im}gSet[M]= & \sum\limits_{I=1}^{N}{(LabSet(X,Y)}+\delta +\max (getIntensity(LabPixSet(I))+ & \,\frac{(Max(LabPiSet(I)}{\left( \frac{G(X,Y)-\theta i}{2} \right)\,}-Th \\\end{align}$ (8)
Step-6: The features are extracted from the segmented and denoised image. The features extracted will be used for selection of best features among the feature set. The feature extraction process considers only the relevant portion in the segmented image. The feature extraction process is performed as
$\begin{aligned} \text { FeatSet }[M]=\sum_{I=1}^R & \frac{\max (\text { getattr }(\text { DenoisedImgSet }(I))}{\operatorname{size}(\text { DenoisedImgSet }(I))} & +\operatorname{corr}(I, I+1)& +\max (\text { getIntensity }(I)) & -\min (\text { getIntensity }(I))+T h\end{aligned}$ (9)
Here getattr() is used for considering the attributes with maximum values excluding the minimum attribute set. The coor() model is used to find the correlation of all features.
Step-7: Display the denoised image set for feature extraction and tumor detection.
}
With the correct blend of reconstruction algorithms and kernels, high-quality CT images were obtained. A signal processing step is used to transform the analogue signals into digital signals, which are then supplied to the digital system for CT image reconstructions. The subject of segmentation of medical images despite the existence of numerous algorithms, remains a complex and demanding one. Classifying segmentation approaches has been attempted by a variety of researchers. In this research edge based segmentation approaches based on grey level and textural feature based methods in the diagnostic image processing field is applied.
Cropping the photos to eliminate black borders was done before to using CNNs to segment the tumour. Cropping was done individually for each of the retrieved segments because of the variations in their dimensions and forms. A sample-wise image standardization method was used to normalize images prior to CNN-based segmentation. It was necessary to enrich the datasets to prevent over fitting and boost the model's capacity to better generalize because of their modest size. This meant that at each cross-validation, the training set was increased by a factor of ten and random transformations were applied to each frame, including image rotation, shift, zoom, horizontal and vertical flip, and horizontal and vertical flip. More precise extraction of target, cell area, perimeter and other multi dimensional features is achieved through the use of edge based segmentation method using CNN. The data set is considered from the link https://data.mendeley.com/datasets/ftmp4cvtmb/1. This research proposes a Precise Denoising Model with Edge based Segmentation for Labeled Pixel Extraction with Fixed Feature Set (PDM-ES-LPE-FFS) model for relevant feature extraction. The proposed model is compared with the Partial Differential Equation Image Segmentation Model (PDEISM) and the results represent that the proposed model performance is better.
There are a total of 1224 MRI pictures available here. Each group of images is presented in two distinct resolutions. The first group consists of 89 normal oral epithelium histopathology images and 439 high-magnification images of oral squamous cell carcinoma (OSCC). The second group includes 201 photos of healthy oral epithelium and 495 histological photos of OSCC at 400x magnification. Tissue slides from 230 patients were collected, processed, and catalogued by medical professionals before being imaged using a Leica ICC50 HD microscope.
Enhanced images are used in many different areas, including medical imaging, where the goal is to restore lost or damaged features. With the advent of CNN, the field of picture enhancement has advanced quickly. Image enhancement is the process of making an image more understandable to humans and more useful as input for other automated image processing methods. The primary goal of image enhancement is to adjust image properties such that they are more suited to a specific job and observer for pixel extraction. The image quality enhancement accuracy levels of the proposed and traditional models are shown in Figure 4.
Reporting the fraction of an image's pixels that were properly labeled is another option for measuring the efficacy of a semantic segmentation. It is usual practice to report pixel accuracy both by class and as a whole. Combining similar portions or segments of an image over their corresponding class labels is the goal of image segmentation, a sub-domain of machine learning and digital image processing. With image segmentation, classification and localization is performed on top with image classification. As a result, picture segmentation is a further generalization of image classification, with the model identifying the location of the object's boundary to determine if it is present. The Figure 5 represents the Image Segmentation Accuracy Levels of the existing and proposed models.
Figure 4. Image quality enhancement accuracy levels
Figure 5. Image segmentation accuracy levels
Figure 6. Image edge segmentation accuracy levels
In image processing, edge detection refers to the process of identifying the edges of visible objects in a CT image. Brightness changes are what trigger its action. Data extraction and segmentation from images rely on edge detection in fields including image processing, machine learning, and machine vision. An integral aspect of segmenting an image is detecting its edges. Many tasks in image processing and computer vision rely on accurately identifying relevant edges. It is a method for finding breaks in intensity in a digital picture. The algorithms used to segment images typically rely on the differences and similarities in the intensity levels included in the input images. Partitioning a picture into sections that are comparable according to a set of established criteria, as opposed to the discontinuity technique, which partitions an image based on sharp changes in intensity, is known as a similarity approach. The image edge segmentation accuracy levels of the existing and the proposed models are shown in Figure 6.
The size of a pixel in an image is determined by the screen's resolution. The image pixel extraction time levels of the proposed model are very low when compared to the existing model that is represented in Figure 7.
Figure 7. Image pixel extraction time levels
There can only ever be one label assigned to a given pixel. In order to train semantic segmentation algorithms, users need to have access to ground truth data, which is where the labels come in. Pixel labelling refers to the action of locating and noting specific features within an image. When producing meta data for disease detection based on image content automatically, pixel labelling is a helpful process. The pixel labeling accuracy levels of the proposed and existing models are shown in Figure 8.
Figure 8. Pixel labeling accuracy levels
The goal of picture denoising is to recover an original image from one that has been corrupted by noise. Denoising can help improve image quality, but since noise, edge, and texture are all high frequency components, it can be tricky to tell them apart. Noise has a negative impact on other image processing activities that may follow, including video processing, picture analysis, and tracking. Image denoising helps to accurately remove the noise that is used for feature extraction that helps in oral cancer detection. The image denoising accuracy levels of the existing and proposed models are shown in Figure 9.
Figure 9. Image denoising accuracy levels
Feature extraction is used to describe the method by which raw data from images are converted into a set of quantifiable features that can be further processed without losing any of the information contained in the original data set for CNN based training. The outcomes are superior to those achieved by simply applying CNN to the data. Feature extraction is useful for cleaning up large data sets by eliminating unnecessary information. When all is said and done, data reduction makes it easier for machines to construct the model and boosts the speed with which CNN learning processes both learn and generalise. The Feature extraction accuracy levels of the proposed and traditional models are shown in Figure 10.
Figure 10. Feature extraction accuracy levels
Noise can be injected into an image throughout the acquisition and transmission process. Several variables can contribute to the appearance of noise in an image. The number of errant pixels in an image is used to estimate the amount of noise present. Invisible specks on a digital photograph taken under ideal lighting conditions to optical and radio astronomy photos that are virtually entirely noise, advanced processing can recover a little bit of information. A photograph with this much noise would be useless since it would be difficult to tell what was being photographed. The suggested caries detection approach presented in this thesis makes use of image processing to identify dental caries in radiographs. The final outcomes are highly dependent on the image quality when it reaches the diagnostic technique. Next, we'll look at how to add weight regularisation methods during segmentation to improve the final result. In order to remove noise from the dataset used, only minimal preprocessing was required. Adding more images to the dataset through various processing methods will benefit the model in its deep learning because the dataset only has a handful to work with. Dental caries segmentation and detection may improve in the future, according to a number of promising research directions. Real image denoising has received little attention despite the numerous in-depth studies on the removal of AWGN. It is difficult to deal with genuine noises because AWGN is so much simpler. An in-depth assessment of a denoiser is difficult in this scenario. White balance, colour decomposition, noise reduction, colour transformation, and compression are all part of the camera's in-camera processing pipeline. External and internal factors, including as lighting, CCD/CMOS sensors, and camera shake, all affect the quality of the final image. Denoising is a difficult problem to solve, even though deep learning is improving rapidly. The fundamental cause for this is a paucity of image pairs for training in real-world denoising methods. As far as we know, all existing denoising methods are taught using data created by adding AWGN to clean photos. However, for real-world denoising, we found that the CNNs trained on simulated data were ineffective. To summarise, the purpose of this study is to provide an overview of the denoising algorithms currently accessible. In order to come up with new denoising schemes, it is helpful to analyse noise in order to come up with different denoising approaches. First, we need to figure out how to deal with noises that occur in the real world. Secondly, it is still a challenge to train deep models without the use of image pairs.
[1] Tian, C., Chen, Y. (2019). Image segmentation and denoising algorithm based on partial differential equations. IEEE Sensors Journal, 20(20): 11935-11942. https://doi.org/10.1109/JSEN.2019.2959704
[2] Manimala, M.V.R., Dhanunjaya Naidu, C., Giri Prasad, M.N. (2021). Sparse MR image reconstruction considering Rician noise models: A CNN approach. Wireless Personal Communications, 116(1): 491-511. https://doi.org/10.1007/s11277-020-07725-0
[3] Tian, C., Xu, Y., Zuo, W. (2020). Image denoising using deep CNN with batch renormalization. Neural Networks, 121: 461-473. https://doi.org/10.1016/j.neunet.2019.08.022
[4] Manimala, M.V.R., Naidu, C.D., Prasad, M.G. (2020). Convolutional neural network for sparse reconstruction of MR images interposed with gaussian noise. Journal of Circuits, Systems and Computers, 29(7): 2050116. https://doi.org/10.1142/S0218126620501169
[5] Chaudhary, A., Bhattacharjee, V. (2020). An efficient method for brain tumor detection and categorization using MRI images by K-means clustering & DWT. International Journal of Information Technology, 12: 141-148. https://doi.org/10.1007/s41870-018-0255-4
[6] Kim, G., Cho, J., Kang, M. (2020). Cauchy noise removal by weighted nuclear norm minimization. Journal of Scientific Computing, 83: 1-21. https://doi.org/10.1007/s10915-020-01203-2
[7] Singh, M., Pandove, G. (2020). Design and simulate different digital image enhancement techniques for biomedical applications using VIVADO system generator. International Journal of Information Technology, 12: 397-401. https://doi.org/10.1007/s41870-019-00295-w
[8] Kalaiselvi, T., Padmapriya, S.T., Sriramakrishnan, P., Somasundaram, K. (2020). Deriving tumor detection models using convolutional neural networks from MRI of human brain scans. International Journal of Information Technology, 12: 403-408. https://doi.org/10.1007/s41870-020-00438-4
[9] Goyal, B., Dogra, A., Agrawal, S., Sohi, B.S., Sharma, A. (2020). Image denoising review: From classical to state-of-the-art approaches. Information Fusion, 55: 220-244. https://doi.org/10.1016/j.inffus.2019.09.003
[10] Nasrin, S., Alom, M.Z., Burada, R., Taha, T.M., Asari, V.K. (2019). Medical image denoising with recurrent residual u-net (r2u-net) base auto-encoder. In 2019 IEEE National Aerospace and Electronics Conference (NAECON), Dayton, OH, USA, pp. 345-350. https://doi.org/10.1109/NAECON46414.2019.9057834
[11] Mittal, K., Aggarwal, G., Mahajan, P. (2019). Performance study of K-nearest neighbor classifier and K-means clustering for predicting the diagnostic accuracy. International Journal of Information Technology, 11: 535-540. https://doi.org/10.1007/s41870-018-0233-x
[12] Lundervold, A.S., Lundervold, A. (2019). An overview of deep learning in medical imaging focusing on MRI. Zeitschrift für Medizinische Physik, 29(2): 102-127. https://doi.org/10.1016/j.zemedi.2018.11.002
[13] Schlemper, J., Yang, G., Ferreira, P., Scott, A., McGill, L.A., Khalique, Z., Gorodezky, M., Roehl, M., Keegan, J., Pennell, D., Firmin, D., Rueckert, D. (2018). Stochastic deep compressive sensing for the reconstruction of diffusion tensor cardiac MRI. In: Frangi, A., Schnabel, J., Davatzikos, C., Alberola-López, C., Fichtinger, G. (eds) Medical Image Computing and Computer Assisted Intervention – MICCAI 2018. MICCAI 2018. Lecture Notes in Computer Science(), vol. 11070. Springer, Cham. https://doi.org/10.1007/978-3-030-00928-1_34
[14] Seitzer, M., Yang, G., Schlemper, J., Oktay, O., Würfl, T., Christlein, V., Wong, T., Mohiaddin, R., Firmin, D., Keegan, J., Rueckert, D., Maier, A. (2018). Adversarial and perceptual refinement for compressed sensing MRI reconstruction. In: Frangi, A., Schnabel, J., Davatzikos, C., Alberola-López, C., Fichtinger, G. (eds) Medical Image Computing and Computer Assisted Intervention – MICCAI 2018. MICCAI 2018. Lecture Notes in Computer Science(), vol. 11070. Springer, Cham. https://doi.org/10.1007/978-3-030-00928-1_27
[15] Han, Y., Yoo, J., Kim, H.H., Shin, H.J., Sung, K., Ye, J.C. (2018). Deep learning with domain adaptation for accelerated projection-reconstruction MR. Magnetic Resonance in Medicine, 80(3): 1189-1205. https://doi.org/10.1002/mrm.27106
[16] Hyun, C.M., Kim, H.P., Lee, S.M., Lee, S., Seo, J.K. (2018). Deep learning for undersampled MRI reconstruction. Physics in Medicine & Biology, 63(13): 135007. https://doi.org/10.1088/1361-6560/aac71a
[17] Meiniel, W., Olivo-Marin, J.C., Angelini, E.D. (2018). Denoising of microscopy images: A review of the state-of-the-art, and a new sparsity-based method. IEEE Transactions on Image Processing, 27(8): 3842-3856. https://doi.org/10.1109/TIP.2018.2819821
[18] Gökdağ, Y.E., Şansal, F., Gökdel, Y.D. (2019). Image denoising using 2-D wavelet algorithm for Gaussian-corrupted confocal microscopy images. Biomedical Signal Processing and Control, 54: 101594. https://doi.org/10.1016/j.bspc.2019.101594
[19] Yu, H., Ding, M., Zhang, X., Wu, J. (2018). PCANet based nonlocal means method for speckle noise removal in ultrasound images. PloS One, 13(10): e0205390. https://doi.org/10.1371/journal.pone.0205390
[20] Fan, L., Zhang, F., Fan, H., Zhang, C. (2019). Brief review of image denoising techniques. Visual Computing for Industry, Biomedicine, and Art, 2(1): 1-12. https://doi.org/10.1186/s42492-019-0016-7
[21] Mafi, M., Martin, H., Cabrerizo, M., Andrian, J., Barreto, A., Adjouadi, M. (2019). A comprehensive survey on impulse and Gaussian denoising filters for digital images. Signal Processing, 157: 236-260. https://doi.org/10.1016/j.sigpro.2018.12.006
[22] Gavaskar, R.G., Chaudhury, K.N. (2018). Fast adaptive bilateral filtering. IEEE Transactions on Image Processing, 28(2): 779-790. https://doi.org/10.1109/TIP.2018.2871597
[23] Bae, W., Yoo, J., Chul Ye, J. (2017). Beyond deep residual learning for image restoration: Persistent homology-guided manifold simplification. 2017 IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), Honolulu, HI, USA, pp. 1141-1149. https://doi.org/10.1109/CVPRW.2017.152
[24] Isogawa, K., Ida, T., Shiodera, T., Takeguchi, T. (2017). Deep shrinkage convolutional neural network for adaptive noise reduction. IEEE Signal Processing Letters, 25(2): 224-228. https://doi.org/10.1109/LSP.2017.2782270
[25] Tian, C., Xu, Y., Fei, L., Wang, J., Wen, J., Luo, N. (2019). Enhanced CNN for image denoising. CAAI Transactions on Intelligence Technology, 4(1): 17-23. https://doi.org/10.1049/trit.2018.1054
[26] Park, B., Yu, S., Jeong, J. (2019). Densely connected hierarchical network for image denoising. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), Long Beach, CA, USA, pp. 2104-2113. https://doi.org/10.1109/CVPRW.2019.00263
[27] Liu, P., Zhang, H., Lian, W., Zuo, W. (2019). Multi-level wavelet convolutional neural networks. IEEE Access, 7: 74973-74985. https://doi.org/10.1109/ACCESS.2019.2921451
[28] Yu, S., Park, B., Jeong, J. (2019). Deep iterative down-up CNN for image denoising. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), Long Beach, CA, USA, 2019, pp. 2095-2103 https://doi.org/10.1109/CVPRW.2019.00262
