Grzegorz Surówka

Different Learning Paradigms for the Classification of Melanoid Skin Lesions Using Wavelets

We use the wavelet-based decomposition to generate the multiresolution representation of dermatoscopic images of potentially malignant pigmented lesions. Three different machine learning methods are experimentally applied, namely neural networks, support vector machines, and Attributional Calculus. The obtained results confirm that neighborhood properties of pixels in dermatoscopic images are a sensitive probe of the melanoma progression and together with the selected machine learning methods may be an important diagnostic tool.

Computational Intelligence and Image Processing Methods for Applications in Skin Cancer Diagnosis

Digital photography provides new powerful diagnostic tools in dermatology. Dermoscopy is a special photography technique which enables taking photos of skin lesions in chosen lighting conditions. Digital photography allows for seeing details of the skin changes under various enlargements and coloring. Computer-assisted techniques and image processing methods can be further used for image enhancement and analysis and for feature extraction and pattern recognition in the selected images. Special techniques used in skin-image processing are discussed in detail. Feature extraction methods and automated classification techniques based on statistical learning and model ensembling techniques provide very powerful tools which can assist the doctors in taking decisions. Performance of classifiers will be discussed in specific case of melanoma cancer diagnosis. The techniques have been tested on a large data set of images.

Semantic analysis of skin lesions using radial basis function neural networks

Artificial Neural Networks have been successfully applied to abroad spectrum of complex analysis problems. Computational intelligence is finding more and more applications in computer aided diagnostics, helping doctors to process large quantities of various medical data. In dermatology it is extremely difficult to perform automatic diagnostic differentiation of malignant melanoma based only on dermatoscopic images. Applying artificial intelligence algorithms to explore and search large database of dermatoscopic images allow doctors to semantically filter out image with specified characteristics. This paper presents an approach for characteristic objects classification found in image database of pigment skin lesions, based on radial basis function kernel for artificial neural networks.

Supervised learning of melanocytic skin lesion images

We use MLP and SVM supervised learning methods to discover patterns in the pigmented skin lesion images. This methodology can be treated as a non-invasive approach to early diagnosis of melanoma. Our feature set is composed of wavelet-based multi-resolution filters of the dermoscopy images. Feature selection is done by the Ridge linear models. Discriminating malicious from benign lesion images with the selected classifiers has sensitivity of 89.2-94.7% and specificity of 85-95%.

Symbolic learning supporting early diagnosis of melanoma

We present a classification analysis of the pigmented skin lesion images taken in white light based on the inductive learning methods by Michalski (AQ). Those methods are developed for a computer system supporting the decision making process for early diagnosis of melanoma. Symbolic (machine) learning methods used in our study are tested on two types of features extracted from pigmented lesion images: coloristic/geometric features, and wavelet-based features. Classification performance with the wavelet features, although achieved with simple rules, is very high. Symbolic learning applied to our skin lesion data seems to outperform other classical machine learning methods, and is more comprehensive both in understanding, and in application of further improvements.

Inductive learning of skin lesion images for early diagnosis of melanoma

We take advantage of natural induction methods to build classifiers of the pigmented skin lesion images. This methodology can be treated as a non-invasive approach to early diagnosis of melanoma. We use the AQ21 application, which is based on the attributional calculus, to discover patterns in the skin images. Our classifier has good efficiency and may potentially be an important diagnostic aid.

On Optimal Wavelet Bases for Classification of Melanoma Images Through Ensemble Learning

This article addresses the medical problem of early detection of the malignant melanoma skin cancer. We present ensemble classification of dermoscopic skin lesion images into two classes: malignant melanoma and dysplastic nevus. The features used for classification are derived from wavelet decomposition coefficients of the image. Our research purpose is to select the best wavelet bases in terms of AUC classification performance of the ensemble. The ensemble learning is optimized by some common quality measures: accuracy, precision, F1-score, FP- rate, specificity, BER and recall. Within the statistics of our machine learning experiments the best model of melanoma uses reverse bi-orthogonal wavelet pair (3.1) and is optimized by FP-rate. This wavelet base performs very well with downscaled image resolutions which matters future small ARM-based devices for computer aided diagnosis of melanoma.

Resolution Invariant Neural Classifiers for Dermoscopy Images of Melanoma

This article contributes to the Computer Aided Diagnosis (CAD) of melanoma pigmented skin cancer. We test back-propagated Artificial Neural Network (ANN) classifiers for discrimination in benign and malignant skin lesions. Features used for the classification are derived from wavelet decomposition coefficients of the dermoscopy image. We show the most efficient ANN setups as a function of the structure of hidden layers and the network learning algorithms. Our network topologies are limited for the sake of restrictions in memory and processing power of smartphones which are more and more popular as hand-held ‘mobile’ CAD devices for melanoma. We claim resolution invariance of the selected wavelet features.

On optimal wavelet bases for classification of skin lesion images through ensemble learning

In order to recognize early symptoms of melanoma, the fatal cancer of the skin, systems for computer aided melanoma diagnosis have been developed for years. In this work we analyze an ensemble-based binary classifier for discriminating melanoma from dysplastic nevus utilizing wavelet-based features of the dermatoscopic skin lesion images. The multiresolution decomposition of the dermatoscopy images is done through wavelet packets. We search for the optimal wavelet base maximizing the quality of the classifier in terms of AUC (Area Under Curve) for models optimized by some common quality measures: accuracy, precision, Fl-score, FP-rate, specificity, BER and recall. Within the statistics of our experiments reverse bi-orthogonal wavelet rbio 3.1 makes the best wavelet model of melanoma.

Applications of Neural Networks in Semantic Analysis of Skin Cancer Images

Computational intelligence is finding more and more applications in computer aided diagnostics, helping doctors to process large quantities of various medical data [Buronni et al. 2004]. In dermatology it is extremely difficult to perform automatic diagnostic differentiation of malignant melanoma based only on dermatoscopic images. Applying artificial intelligence algorithms to explore and search large database of dermatoscopic images allow doctors to semantically filter out image with specified characteristics. This paper presents an semantic approach for characteristic objects classification found in image database of pigment skin lesions, based on radial basis function kernel for artificial neural networks. Presented approach is divided into few parts: JSEG image segmentation [Deng et al. 2001], feature extraction and classification. Prepared features vector consist of color models parts. For classification Artificial Neural Networks and Support Vector Machines are used and their performance is evaluated and compared. Success rates in both cases are greater than 90%.