Digital hair segmentation using hybrid convolutional and recurrent neural networks architecture

BACKGROUND AND OBJECTIVE Skin melanoma is one of the major health problems in many countries. Dermatologists usually diagnose melanoma by visual inspection of moles. Digital hair removal can provide a non-invasive way to remove hair and hair-like regions as a pre-processing step for skin lesion images. Hair removal has two main steps: hair segmentation and hair gaps inpainting. However, hair segmentation is a challenging task which requires manual tuning of thresholding parameters. Hard-coded threshold leads to over-segmentation (false positives) which in return changes the textural integrity of lesions and or under-segmentation (false negatives) which leaves hair traces and artefacts which affect subsequent diagnosis. Additionally, dermal hair exhibits different characteristics: thin; overlapping; faded; occluded and overlaid on textured lesions. METHODS In this presented paper, we proposed a deep learning approach based on a hybrid network of convolutional and recurrent layers for hair segmentation using weakly labelled data. We utilised the deep encoded features for accurate detection and delineation of hair in skin images. The encoded features are then fed into recurrent neural network layers to encode the spatial dependencies between disjointed patches. Experiments are conducted on a publicly available dataset, called "Towards Melanoma Detection: Challenge". We chose two metrics to evaluate the produced segmentation masks. The first metric is the Jaccard Index which penalises false positives and false negatives. The second metric is the tumour disturb pattern which assesses the overall effect over the lesion texture due to unnecessary inpainting as a result of over segmentation. The qualitative and quantitative evaluations are employed to compare the proposed technique with state-of-the-art methods. RESULTS The proposed approach showed superior segmentation accuracy as demonstrated by a Jaccard Index of 77.8% in comparison to a 66.5% reported by the state-of-the-art method. We also achieved tumour disturb pattern as low as 14% compared to 23% for the state-of-the-art method. CONCLUSION The hybrid architecture for segmentation was able to accurately delineate and segment the hair from the background including lesions and the skin using weakly labelled ground truth for training.