Seven-Point Checklist with Convolutional Neural Networks for Melanoma Diagnosis

Reliable skin lesion detection is an important pre-requisite for melanoma and other skin diseases diagnosis. Existing melanoma assessment models consider either pattern analysis methods or seven-point checklist criteria to investigate skin lesion. However, automatic and accurate detection of the skin lesion and subsequently melanoma diagnosis remain an unresolved challenge. Furthermore, there is limitations in both approaches and a trade-off between the two assessment strategies. This paper proposes a pattern analysis method incorporated with seven-point checklist exploiting convolutional neural network for melanoma diagnosis where the lesion features are extracted automatically. The benefit of features learned automatically from the dermoscopic images through the stacked layers of convolution filters have been designed, realised and evaluated. Both clinical and dermoscopic images have been considered as input to the developed multiple-input convolutional neural networks (CNNs) where a separate feature extraction model is implemented for each image type. The features produced from both models are concatenated for interpretation and final lesion type prediction. The sum of the weight for predicated lesions which are calculated according seven-points check list criteria are then passed into a threshold model to decide whether the image normal or abnormal (melanoma or non-melanoma). Performance of the developed algorithm is assessed using a dataset of 2000 dermoscopic images. The initial results obtained from the proposed system show a convincing and promising ability for lesion detection and automated melanoma diagnosis from dermoscopy images.