Melanoma is the deadliest form of skin cancer, with a high mortality rate every year, and New Zealand is known to be one of the countries with the highest incidence of this disease. Overexposure to ultraviolet sun rays causes the upper layers of the skin to produce a pigment known as melanin, the primary cause of melanoma. An early diagnosis and prognosis of melanoma can improve survival rates before it becomes dangerous. There is a high demand for computerized automated skin lesion analysis techniques at low cost since trained specialists are in limited supply, and manual diagnosis involves high costs. The increasing use of non-invasive methods in diagnosing malignant melanoma reduces the need for biopsies. A primary objective of this thesis is the development of computerized detection systems for analyzing lesions and distinguishing melanoma from other types of skin cancer. The thesis focuses mainly on designing novel methods for three major phases of automatic melanoma diagnosis process: a) Pre-processing deals with removal of noise artefacts such as hairlines and improving image’s contrast, b) Lesion segmentation to accurately extract lesion region using deep learning, and c) Melanoma classification using deep learning for a fast and accurate detection.

For the first phase, an algorithm Intensity Adjustment-based Hair Removal (IA-HR) employing morphological operators is developed to remove the hairlines, which are a significant problem in skin image samples. Additionally, a Multi-scale Context Aggregation Convolutional Neural Network (MCACNN) is used to enhance the contrast and resolution of images. A cleaned dataset is then generated using these pre-processing methods. A class imbalance problem is also addressed using data augmentation methods.

To determine lesion borders and extract lesion information, deep learning networks are used. Two network designs were constructed; one was based on encoder-decoder layered patterns (EDNet) and the other on atrous convolutions (DilatedSkinNet). As DilatedSkinNet shows higher average accuracy, thus it is preferred to EDNet for segmentation tasks.

Finally, in the classification stage, we present the design of a multi-layer deep convolutional neural network (DCNN) named as LCNet to distinguish melanoma from benign tumors. The designed classification network is a lightweight network having a smaller number of learnable parameters. Due to the less complex architecture of classification network, it takes less inference time. Moreover, the network can also diagnose diseases efficiently and accurately without pre-processing and segmentation techniques, unlike traditional machine learning classifiers that rely heavily upon these initial steps.

The study examined, however, the effects of applying designed pre-processing and segmentation methods on segmentation and classification performance. The aim of this effort is to further enhance the performance of the classifier by preparing more rich and clean data for training. The classification network is fed with hairlines-free, high-contrast, and segmented images, and its performance is compared to raw images. Our experiments showed that the classification network efficiently processes raw and complex data by offering an accuracy (ACC) of 90.92±1.0%. The accuracy performance of classification model is improved with pre-processed data as 92.47% and with pre-processed+segmented data as 93.40% indicating the classification model’s performed higher with pre-processing and segmentation operations to distinguish melanoma vs benign. Additionally, it is observed that noise-free and cleaned data using IA-HR and MCACNN methods improved the performance of the segmentation approach.

In our study, we found that the proposed pre-processing and segmentation methods could improve the performance of deep learning-based classifiers. Furthermore, there are many areas of melanoma diagnosis process where the proposed approaches can be successfully applied. The denoising method can be used to clean skin samples without causing any discomfort to the patients such as hairlines can be automatically eliminated from images and contrast can be enhanced. Another use of segmentation model to extract lesion region and to perform detailed analysis of it. Additionally, segmentation method can be employed to generate accurate and smooth ground truth labels for new samples that are currently annotated manually by experts. The classification model can be used to classify melanoma in less time such as in 1.3 seconds as predicted by our model. This method may also be used to generate a handcrafted feature extraction process if a machine learning-based classifier is employed to diagnose skin cancer. The DCNN classification model also showed more improvement in diagnosing melanoma when trained on a large, balanced, and pre-processed dataset. In its future scope, embedded systems such as FPGA based system-n-chip and other resource-constrained implementations can benefit from the designed classification network.