Clinical Validation of an automated system for supporting the early diagnosis of melanoma

ckground: Early diagnosis and surgical excision is the most effective treatment of melanoma. Well-trained dermatologists reach a high level of diagnostic accuracy with good sensitivity and specificity. Their performances increase using some technical aids as digital epiluminescence microscopy. Objective: The purpose of this study is to evaluate a multiple-classifiers system for supporting the early diagnosis of melanoma. The performance of the system was compared to that of a group of eight expert dermatologists and it was also tested as diagnostic support for early melanoma. Methods: MEDS (MElanoma Diagnosis System) is an automatic digital system, which allows dermatologists to acquire a D-ELM image of melanocytic lesions (MLs), and then automatically extracts a set of geometric, morphologic and colorimetric features. MEDS assesses the diagnosis by combining the diagnostic outputs of three different classifiers: linear discriminant analysis, k-nearest neighbor and decision tree. MEDS is trained and validated on a set of 152 MLs. Results: The eight dermatologists have sensitivity and specificity values of 0.83 and 0.66 respectively. None of the single classifiers reaches the clinicians’ values for both the parameters. The combination of the classifier shows that the 3-Classifiers systems perform as well as the eight dermatologists (sensitivity range: 0.75 ÷ 0.86; specificity range: 0.64 ÷ 0.89). The further combination of MEDS with the dermatologists shows an average improvement of 11% (p = 0.022) for what concerns physicians’ sensitivity. Conclusion: MEDS has comparable performance with respect to those of the dermatologists and it improves their sensitivity when used in a supporting mode. This fact suggests that an automated system may be effective in supporting dermatologists in the recognition of early melanomas. Although only 5% of skin cancers are melanomas, this tumor is responsible for 91% of deaths due to skin cancer. Its incidence is increasing worldwide.1 The early diagnosis of melanoma is the principal determinant in its prognosis.2 Diagnosis is difficult and requires a well-trained dermatologist, because the early malignant lesion can have a benign appearance.3,4 Several studies have shown that the diagnostic accuracy of a specialist is about 69% for early melanomas, and it reduces to 12% for non-specialists.5 Digital epi-luminescence microscopy (D-ELM) is one of the techniques that had considerable success in clinical practice (see for a review Zsolt).6 It allows the visualization of several morphological and structural characteristics of skin lesions at the naked eye, providing the physician with additional diagnostic criteria. A standardized procedure to assess the relevant diagnostic features has been established. In particular the ABCD rule of dermatoscopy helps physicians to assess a diagnosis by evaluating some characteristic of a ML.7,8 To avoid any bias in clinical judgment, computer-aided diagnosis systems have been introduced. After the first experiences with SkinView,9,10 several automatic systems were proposed for the early diagnosis of melanoma, using different approaches.11-20 In this work we present a clinical validation of MEDS (MElanoma Diagnosis System), a computer-based system for automatic classification of MLs. It combines digital image acquisition and processing with machine learning techniques. In particular MEDS processes D-ELM images extracting features that could be meaningful for the expert dermatologist, following the so-called ABCD Rule. The features are the input of three different classifiers, namely Linear Discriminant Analysis, Decision Tree and k-Nearest Neighbor, which MEDS integrates by means of voting schemata. The main goal of MEDS is to provide support to physicians for the early diagnosis of melanoma. This study has two main aims: firstly, the comparison of MEDS’ performances with respect to eight dermatologists; secondly, the evaluation of MEDS as an effective tool...