Automatic differentiation of melanoma from dysplastic nevi
Introduction
Malignant melanoma is a type of skin cancer and although it accounts for less than 2% of all skin cancer cases, it is the deadliest type and causes the vast majority of deaths [1]. According to the latest report, melanoma caused over 20,000 deaths annually in Europe [2]. American Cancer Society also reported the estimated deaths of melanoma in 2013 as 9480 individuals and new cases as 76,690 individuals. Nevertheless melanoma is the most treatable kind of cancer emphasizing the importance of early diagnosis. Currently in the clinical field, the “ABCDE” rule is the clinical routine which is used to detect malignant melanoma at its earliest stage [3]. This routine seeks for characteristics synonymous with malignant melanoma at its early stage and can be listed as: asymmetry, irregular borders, variegated colours, diameters greater than 6 mm and evolving stages. The detection is performed by visual inspection and further analysis, using clinical imaging techniques such as dermoscopic imaging. Visual inspection, similarity between different lesions and the necessity to perform patient follow-up over years make the diagnosis task difficult for the dermatologists and more prone to errors notably the repeatability of the detection. Due to this fact and to the importance of early diagnosis of melanoma, the research communities have dedicated their efforts to developing computerized lesion analysis algorithms. These computer-aided systems implement automatic processing to facilitate the task of specialists and focus on different aspects such as segmentation, detection and classification of the lesions. In this work we focus on classification task, particularly discriminating melanoma and dysplastic nevi. It has been reported that from 2 to 8% of the Caucasian population have dysplastic nevi (also known as atypical moles, i.e. unusual benign moles that may resemble melanoma) [4]. Individuals who have dysplastic nevi syndrome or dysplastic nevi with family history of melanoma face a higher risk of developing melanoma. However, only a small number of these dysplastic nevi might develop into melanoma and most dysplastic nevi will never become cancer [5]. Notably discrimination of these two categories is more challenging due to their similarities [6].
In the past decade, different approaches have been proposed for detection of melanoma. The developed methods used the common classification framework in the computer vision field: the pipeline of this framework usually consists of the four following steps: (i) segmentation, (ii) feature detection, (iii) feature selection/extraction and (iv) classification. Furthermore, those methods were based on two different screening imaging techniques, either clinical or dermoscopic images. Clinical imaging was the first screening technique and the initial research focused on this modality. This technique acquires images (digital or not) of skin lesions and represents what a clinician sees with naked eyes [29].
Later, clinical imaging was replaced by a more suitable technique such as dermoscopy imaging (also known as epiluminescence microscopy). This techniques has been used extensively by dermatologists and researchers, and several frameworks based on this modality has been proposed by the research community to detect malignant melanoma [30], [31], [32].
Fig. 1, depicts a summary of most of these methods by reporting their results in terms of sensitivity, specificity and dataset size. Sensitivity or recall (SE) refers to the number of correctly identified cancer cases to the total number of cancer cases in the dataset and specificity or true negative rate (SP) refers to the proportion of negatives or correctly identified non-cancer cases to the total number of non-cancer cases in the dataset. The methods are also categorized based on their differentiation scope, melanoma from benign (M vs B), melanoma from benign and dysplastic (M vs B + D) and melanoma versus dysplastic only (M vs D). The summary of the state of the art methods showed that fewer attention has been paid to the discrimination of melanoma and dysplastic in the past, which we believe is more challenging for both specialists and automated algorithms.
It is difficult to offer a fair comparison among the methods since their performances are reported in the literature using different datasets. These methods also use multiple machine learning approaches to distinguish melanoma lesions, such as AdaBoost (AdB), artificial neural network (ANN), support vector machines (SVM), k-nearest neighbour (kNN) or random forest (RF). However, AdB, SVM and ANN appear to be the most popular methods. The detected features to feed these classification methods can also be categorized. Table 1 summarizes these features with their corresponding references, by using a similar methodology as in [29].
Features can be extracted either from a global or local manner. A global feature is extracted by taking the lesion as a whole, while local features are extracted more densely. A grid can be defined over the lesion and each descriptor is extracted for each small portion of the grid. The use of local features increases the size of the feature vector. It also increases its computation cost and the complexity of the feature space. Hence, the bag of features (BoF) approach was commonly used to tackle these drawbacks [14], [17], [37], [13], [22], [12], [38].
In this work we focus our research on dermoscopic images and propose an automatic framework for detection of melanoma from dysplastic lesions since their distinction, based on their similarities, is more challenging for both the classifier algorithm, the general practitioner and even the expert dermatologists and fewer studies have addressed the issues (see Fig. 1). We evaluate the performance of well-known texture descriptors besides common colour and shape features to discriminate the two classes. To the best of our knowledge although a combination of colour, shape and texture features were used previously [25], [23], [21], these features were mostly designed to mimic the characteristics of “ABCD” rule. In this research we consider those features besides well-known texture features such as local binary pattern, grey-level co-occurrence matrix, Gabor filter and histogram of gradients and present a comparison between their individual and combined performances, while they are extracted by both global and local methods. We also consider three classifiers such as Support Vector Machine, Gradient Boosting and Random Forest.
The rest of the paper is organized as follows: Section 2 describes the proposed framework. Section 3 illustrates the experiments and proposed validation. The results and discussion are illustrated in Section 4 and finally, the conclusions are presented in Section 5.
Section snippets
Proposed framework
Different sections of our pipelines as illustrated in Fig. 2 are explained in the following.
Image dataset
The experiments have been carried out using a fraction of the dermoscopy dataset of the Vienna General Hospital. This is a large-scale dataset with over 5000 lesions including 4277 benign lesions, 1002 dysplastic lesions and 101 malignant melanoma images. In this dataset the dysplastic and melanoma lesions were surgically removed and their ground truth were provided by histological diagnosis [20]. This dataset was used as well in [58], [39]. Using the proposed segmentation approach we were able
Results and discussion
The obtained results for global-features are illustrated in Table 3. For each feature set the average performance of each classifier is presented over 10 sets, in terms of SE and SP. In the case of melanoma or in general cancer-case classification it is very important to correctly classify the cancer cases i.e. high sensitivity or recall is required. For this reason high sensitivity is considered as the priority measure. In this table, the individual features are sorted with reference to their
Conclusion
In this work we proposed an automatic framework for differentiation of melanoma lesions from dysplastic nevi. These classes were chosen since their distinction, due to their similarities, is a challenging task for both classifiers and specialists. The framework consists of automatic segmentation, feature extraction, feature modification and classification stage. Our proposed segmentation algorithm were able to segment 95.43% of the lesions, while the rest of the images were excluded due to
Acknowledgement
We would like to acknowledge, Dr. Harald Kittler, Dr. Michael Binder and Dr. Harald Ganster, who kindly provided the dermoscopy dataset of Vienna General Hospital for this research [20].
References (63)
- et al.
Atypical mole syndrome and dysplastic nevi: identification of populations at risk for developing melanoma – review article
Clinics (Sao Paulo)
(2011) - et al.
A multiple classifier system for early melanoma diagnosis
Artif Intell Med
(2003) - et al.
A methodological approach to the classification of dermoscopy images
Comput Med Imaging Graph
(2007) - et al.
An improved internet-based melanoma screening system with dermatologist-like tumor area extraction algorithm
Comput Med Imaging Graph
(2008) - et al.
Toward a combined tool to assist dermatologists in melanoma detection from dermoscopic images of pigmented skin lesions
Pattern Recognit Lett
(2011) - et al.
A comparison of machine learning methods for the diagnosis of pigmented skin lesions
J Biomed Inform
(2001) - et al.
Computerized analysis of pigmented skin lesions: a review
Artif Intell Med.
(2012) - et al.
Advances in skin cancer image analysis
Comput Med Imaging Graph
(2011) - et al.
A comparative study of efficient initialization methods for the k-means clustering algorithm
Expert Syst Appl
(2013) - et al.
Feature normalization and likelihood-based similarity measures for image retrieval
Pattern Recognit Lett
(2001)
A novel method for color correction in epiluminescence microscopy
Comput Med Imaging Graph
Automated color calibration method for dermoscopy images
Comput Med Imaging Graph
Colour and contrast enhancement for improved skin lesion segmentation
Comput Med Imaging Graph
Cancer facts and figures 2014
Melanoma incidence and mortality in Europe: new estimates, persistent disparities
Br J Dermatol
Early diagnosis of cutaneous melanoma: revisiting the ABCD criteria
JAMA
Dysplastic naevus vs. in situ melanoma: digital dermoscopy analysis
Br J Dermatol
An ensemble classification approach for melanoma diagnosis
Memet Comput
Tuning cost-sensitive boosting and its application to melanoma diagnosis
Multiple classifier systems
Automated diagnosis of skin cancer using digital image processing and mixture-of-experts
Biomedizinische Technik/Biomed Eng
Automated diagnosis of pigmented skin lesions
Int J Cancer
Modeling spatial relation in skin lesion images by the graph walk kernel
Automatic learning of spatial patterns for diagnosis of skin lesions
Two systems for the detection of melanomas in dermoscopy images using texture and color features
IEEE Syst J
What is the role of color symmetry in the detection of melanomas?
Advances in Visual Computing
On the role of shape in the detection of melanomas
The role of keypoint sampling on the classification of melanomas in dermoscopy images using bag-of-features
Pattern recognition and image analysis
A support vector machine for decision support in melanoma recognition
Exp Dermatol
Four-class classification of skin lesions with task decomposition strategy
IEEE Trans Biomed Eng
Automated melanoma recognition
IEEE Trans Med Imaging
Cited by (87)
Fusion between an Algorithm Based on the Characterization of Melanocytic Lesions' Asymmetry with an Ensemble of Convolutional Neural Networks for Melanoma Detection
2024, Journal of Investigative DermatologyDifferentiating between early melanomas and melanocytic nevi: A state-of-the-art review
2023, Pathology Research and PracticeA survey, review, and future trends of skin lesion segmentation and classification
2023, Computers in Biology and MedicineA comparative analysis of melanoma detection methods based on computer aided diagnose system
2022, Materials Today: ProceedingsCharacterizing the role of dermatologists in developing artificial intelligence for assessment of skin cancer
2021, Journal of the American Academy of DermatologySkin disease diagnosis with deep learning: A review
2021, Neurocomputing