Analyzing the Impact of Image Denoising and Segmentation on Melanoma Classification Using Convolutional Neural Networks

Early skin cancer detection and its treatment are crucial for reducing death rates worldwide. Deep learning techniques have been used successfully to develop an automatic lesion detection system. This study explores the impact of pre-processing steps such as data augmentation, contrast enhancement, and segmentation on improving the convolutional neural network (CNN) performance for lesion classification. The classification network was designed from scratch by uniquely organizing its layers and using a different number of kernels, depth of the network, size, and hyperparameters. In addition, the network’s performance was improved by pre-processing and segmentation steps. The proposed network was compared with the current state-of-the-art to demonstrate its best performance on the benchmark HAM10000 lesion dataset. The experimental study revealed that the classification network using denoised+segmented data achieved an accuracy (ACC), precision (PRE), recall (REC), specificity (SPE), and F-score of 93.40%, 93.45%, 94.51%, 92.08%, and 93.98%, respectively. To conclude, classification performance can be improved by incorporating pre-processing and segmentation steps.

I. INTRODUCTION

The mortality rate from melanoma has continuously increased over three decades, reporting four to five deaths per 100,000 population [1]. Dermoscopy is the primary tool to diagnose lesions, resulting in limited diagnostic performance due to the visual complexity of different types of lesions and requiring expert input. There are primary two types of skin lesions: Benign (further consisting of nevi, basal, and squamous) and Melanoma. All of them present high intraclass variations, such as their color, texture, shape, and size, making it difficult to distinguish between them. The criteria named ABCDE [2] was introduced to assess the characteristics of lesion type; however, an expert requires intensive knowledge to analyze these properties. The human performance for detecting melanoma is less than 80% [3], with a mean sensitivity of 40% [4], and even less in the case of more complex lesion patterns. Additionally, the traditional clinical diagnostic approaches are expensive, prolonged diagnosis timeframes, and are unavailable to many experts [5]. The mortality rate and complications can be drastically reduced if the disease is detected early. In recent years, traditional machine learning techniques (supervised and unsupervised classification algorithms) [6] achieved remarkable performance in different areas, which are extended to the medical field. In addition, deep learning models represented by CNN gained great success by offering a high detection rate for complex data sets. In the current state of deep learning research towards melanoma classification, there have been no comparative analyses of deep learning performance with raw and processed data to identify whether deep learning models are more effective with cleaned and segmented data.