Skin diseases are most common in day to day life of people. Millions of people in the world are suffering from different skin disorders. High-level of expertise needed to diagnose skin diseases because of their visual aspects in variety. To achieve a more reliable and objective diagnosis, a computer aided diagnostic system is considered as the human judgment is subjective and not easily reproducible. In the proposed system, the feasibility of constructing a universal skin disease diagnosis system using deep convolutional neural network(CNN) is investigated.
We train the CNN architecture using the 23,000 skin disease images from the dataset and test its performance with those data, another skin disease dataset, and images. The proposed system can able to achieve as high as 73% Top one accuracy and 91% Top five accuracy when testing on the dataset. It is possible to improve these accuracies further if more training images are used.
Skin diseases are more common infection seen among people. Due to severe disfigurement and related hardships, skin disorders gives much trouble to the affected people.
skin cancer become more serious, in terms of facts and figures and is the most common form of cancer in United States. According to statistics study in year 2012, over 5.4 million cases of non melanoma skin cancer, including basal cell carcinoma and squamous cell carcinoma, are treated in America. Compared to the incidences of cancers of the lung, breast, prostate and colon , number of cases of skin cancer is more every year. Research also shows that 1/5th of Americans will develop a skin cancer during the course of a lifetime,
However, the diagnosis of skin disease is challenging.
To diagnose a skin disease, a variety of visual clues may be used such as the individual lesional morphology, the body site distribution, color, scaling and arrangement of lesions. When the individual components are analyzed separately, the recognition process can be quite complex. For example, the well studied skin cancer, melanoma, has four four major clinical diagnosis methods: ABCD rules, pattern analysis, Menzies method and 7-Point Checklist. To use these methods and achieve a good diagnostic accuracy, a high level of expertise is required as the differentiation of skin lesions need a great deal of experience.
Melanoma is considered the deadliest form of skin cancer. This is because 75% of deaths associated with melanoma skin cancer. In 2013, it is estimated that 76,690 people will be diagnosed with melanoma and 9,480 people will die of melanoma in the United States. According to the statistical data from the World Health Organization (WHO), between 2 and 3 million non-melanoma skin cancers and 132,000 melanoma skin cancers occur globally each year. Due to recognizing the increasing in skin cancer, modern medical science is seeking to assist the dermatologists in their diagnosis without the need for special or expensive equipment. This model will help remote patient with fast and accurate diagnosis of their skin case.
The proposed model works by acquiring skin image, then apply some preprocessing techniques to improve the quality of the image and removing the skin parameters which may affect the skin images and will end with missing diagnosis. Improving the image quality by applying one of the morphological technique for thick hair removal. Then use a median filter to remove structure elements and very thin hair. Finally, the image is ready for segment the infected lesion from the skin image. Segmentation is one of the most important phases, as the resulting segmented lesion is used as an input to feature extraction and disease classification phases. Existing well known feature including shape, color and texture sets combine many ad-hoc calculations and are unable to easily provide intuitive diagnostic reasoning. This paper presents the design and evaluation of a set of features for objectively detecting melanoma in an intuitive and accurate manner by applying deep feature extraction using pre-trained CNN. The last phase to classify skin infected lesion based on CNN features to train a multiclass SVM classifier.
Abcd rule: Finding melanoma early is the key to curing this terrible disease. That is why learning the ABCDE rule for skin cancer is so important. This system provides and easy way to recognize moles and growths that might be cancerous.
Although most of your ‘suspicious’ moles will turn out to be normal non-cancerous moles, it is much better to be safe than sorry. To not see, or simply ignore an early melanoma can be devastating. Because melanoma can disguise itself as a strange looking mole, be sure to review the ABCDE rule for skin cancer to properly identify abnormal growths. If your mole or growth has one or more of the ABCDE’s, you should show it to your doctor as soon as possible!
Disadvantages: In addition to that, some lesions have irregular boundaries and in some cases there is a smooth transition between the lesion and the skin. In automated diagnosis of skin lesions, feature design is based on the ABCD rule of dermatoscopy. ABCD represent the asymmetry, border structure, variegated color, and dermatoscopical structures and define the basis for a diagnosis by dermatologist.
CAD Systems: CAD systems are intended to help dermatologists in different stages of skin lesions analysis from the preview to to classification. The classification stage of a CAD system is responsible for the production of a diagnostic features extracted from the previous block. There are many classification techniques used today, capable of learning from a set of known decisions (training set). Any supervised classification method is a potential candidate, for example, K-Nearest Neighbours method (KNN), Bayes Classifier, Fisher Discriminant, Support Vector Machines (SVM), Boosting Methods, Neural Networks, Decision Trees, amongst others.
The proposed model seeks to design and develop a computer vision based system for segmentation and classification of skin lesions along with the extraction of discriminating set of features from skin lesions for efficient classification.
There are many challenges in dealing with the skin digital images that taken from a normal camera as it contains noises such as hair and air bubbles. These noises may lead to inaccuracy of the classification and model will give the wrong predication result. In order to avoid that, images are exposed to various image processing techniques.
Preprocessing is done to remove the background noises such as hair and air bubbles and other noises in the skin image. First use simple morphological closing operation with a disk-shaped structuring element. Based on the assumption that hair segments are thin structures, a simple morphological technique is applied; next, a hair mask is retained by using a global automatically threshold over the image intensity information. Each hair pixel from the resulted mask is replaced by an average mean of the neighbor’s pixels.
Noise elimination and image smoothing done by using median filtering. Median filtering is used for reducing the effect of small organizations like thin hairs and air bubbles and finally ends up with a smooth image.
Segmentation is one of the most important tasks in image processing and machine vision. Image segmentation methods can be broadly classified into three categories: Edge-based methods, Region-based methods and Pixel-based direct classification methods. In the proposed system, the pixel-based direct classification method is used for segmentation. This involves three major steps: convert color space, feature extraction and clustering.
Convert to HSV color space
Red, Green, Blue or RGB color space is not preferred for color based detection and color analysis because of mixing of color (chrominance) and intensity (luminance) information and its non uniform characteristics. Hue, Saturation, Value or HSV is a color space that describes colors (hue or tint) in terms of their shade (saturation or amount of gray) and their brightness (value or luminance). HSV color space representing color in a way similar to how human perception. The HSV color space provides an intuitive representation and approximates the way in which human perceive. As HSV color space composed of Hue; represents color tone (e.g.: red, pink and blue), Saturation; represents the amount of color (e.g.: bright) and Value; represents the amount of light (e.g.: dark and light). Fig. 4 show the effect of converting RGB skin images to HSV skin image and the corresponding H, S, and V channels.
Feature extraction is one of the most important phases of image processing which requires extensive domain knowledge to help in classification phase. Deep Convolutional Neural Networks (CNNs), a specific type of deep learning algorithm, overcome the problem in traditional machine learning algorithms which required manual feature extraction before the classification process. CNNs not only perform classification, but they can also learn to extract features directly from raw images.
In this paper, an enhanced robust model has been proposed for skin diagnosis using skin lesion image obtained from a standard camera. Using CNN as representative and discriminative feature extractor allows the model to represent its diagnosis in the effective solution for automated recognition of skin diseases. On one hand, this would be useful for dermatologists to reduce diagnostic errors and help remote patients to diagnosis and their skin lesion at a reduced cost while reducing over dependence on medical experts. Experimental results indicate that CNN features easily outperform hand-engineered features in terms of better sensitivity, specificity, and accuracy. The results of extensive experimental trials revealed that the proposed model produced a significant improvement of around 11% in diagnostic accuracy compared to the best of the other state of the art computer aided skin diagnosis.
In our future research, we hope we can push this work much further and get a better accuracy. There are several places we can improve our work. First, since Image Net data are not specialized for skin data, the CNN model from scratch and test its performance. Second, the Dermnet images are organized using a biological taxonomy which is not the best choice for computer vision applications. We will work with a dermatologist to design a visually organized taxonomy and apply it to our classifier 2. Third, as the experimental results suggest that more variance in the training set would lead to a better accuracy, we should increase the size of our training set. Also note that the images retrieved by the networks are closely related to the ground truth. We may need to design a hierarchical classification algorithm using the retrieved images to improve the accuracy.