C LASSIFICATION OF SKIN LESIONS IN DERMOSCOPY IMAGES Ana Raimundo Jorge Martins Júlia Pinheiro Ricardo Trindade Introdução à Engenharia Biomédica Tutored by Professora Margarida Silveira 13th and 14th December 2011
Dermoscopy Melanoma Existing non-automatic methods ▫Pattern Analysis ▫ABCD rule ▫7-Point checklist Automatic Methods ▫Segmentation ▫Feature extraction ▫Classification Example of an automated method Investigation ▫ADDI Conclusion
Support technique for physicians Non-invasive in vivo method Allows to obtain detailed images of the lesions Crucial in most diagnoses
Malignant poliferation of melanocytes with the potential to metastasize Deadliest type of skin cancer
Increased incidence over the last decades Chance of healing is related to the tumor excision in the early stages of its development
Pattern AnalysisABCD Rule7-Point Checklist
Observation of the characteristics of the lesion Comparison with general benign and melanoma features Unreliable diagnosis
Attribution of a score based on the analysis of four characteristics: ▫Assimetry (score from 0 to 2); ▫Border (score from 0 to 8); ▫Color (score from 1 to 6); ▫Differential structure (score from 1 to 5);
Final score (Total Dermoscopy Score): (A score x 1,3) + (B score x 0,1) + (C score x 0,5) + (D score x 0,5) Total Dermoscopy Score (TDS) Interpretation < 4.75Benign melanocytic lesion Suspicious lesion > 5.45 Highly suspicious for melanoma False positive (> 5.45)Other pigmented lesions
Analysis of 7 characteristics usually found in melanoma (Major and minor criteria) More efficient and simple than Pattern analysis
Major criteria (score of 0 or 2): ▫Atypycal pigment network ▫Blue-whitish veil ▫Atypycal vascular pattern
Minor criteria (score of 0 or 1): ▫Irregular streaks ▫Irregular pigmentation ▫Irregular dots/globules ▫Regression structures
Subjectivity Slow appliance Unreachable to untrained dertamologists Motivation Simple automatic methods Appliable on a larger scale Reliable and objective Objective
Automatic Methods Segmentation Feature extraction Classification
Pre-Processing the image ▫Improvement of the resolution ▫Removal of the hairs
Isolation of the lesion from the rest of the image Makes the analysis of the lesion easier
Based in forming a curve and then deforming it Geodesic curve obtained Unreliable method for subtle boundaries and noisy images Geodesic Active Contours Approach User interaction; Connects points marked by the user on the boundaries of the lesion; Useful in noisy images; Geodesic Edge Tracing
Focused on the brigthness component of the image Selection of different gray areas and comparison with the total value of the image area Adaptive Segmentation of gray areas
Similar to the geodesic methods Adapts the curve to the region of the lesion, instead of the border Level Set Segmentation Finds line segments on the image; Estimates the distance between them and connects them with a curve; Robust Snakes Compares the pixels of the image with a Threshold Classificates each pixel as part of the image or the lesion It changes the Threshold to fit each image Active Thresholding
The simplest ones consist in using formulas to search in the lesion specific indicators of melanoma Probabilistic classifiers extract several features related to melanoma from a data base and then calculate the probability of the lesion being a melanoma
The most complex methods analyze various components: ▫Shape features ▫Color features ▫Texture features
Final step of the diagnosis process Automatically says if the lesion is bening or a possible melanoma to be analised
Several methods like: ▫KNN (K-Nearest Neighbour) ▫SVM (Support Vector Machine)
Segmentation Original Wrong segmentation Final
Segmentation Feature extraction Classification
Relation between primary melanoma and lymph nodes location – Sentinel Nodes Allows phycisans to know which are the nodes to analyse
ADDI (Automated computer-based diagnosis system for dermoscopy images) Partnership between Universidade do Porto, Universidade de Aveiro, Instituto Superior Técnico, Institute of Systems and Robotics and Hospital Pedro Hispano
The most efficient diagnosis method is still the manual one In the present days, there are several efficient automatic methods available The best solution is to alternate between different techniques, considering the objective The analysed algorithms are successfully applied on a larger scale, saving time and reducing costs