IDUNet++: an improved convolutional neural network for melanoma skin lesion segmentation based on UNet++

Abstract

Melanoma originates from the malignant transformation of melanocytes, it can gradually spread and metastasize. As the most aggressive and deadly type of skin cancer, melanoma posed a significant threat to patient health, but early diagnosis and intervention can improve patient survival and improve the prognosis of poor patients. Computer-aided diagnosis can help dermatologists to make early diagnoses of melanoma. UNet++ as a more advanced model among existing segmentation algorithms has practical value in the segmentation and diagnosis of melanoma, but after experiments, we found that its segmentation performance still has much room for improvement. In the study, we tried to improve the model performance based on the UNet++ algorithm, and a new convolutional neural network IDUNet++ (Inception Dilated UNet++) for melanoma skin lesion segmentation by introducing Inception block and dilated convolution was proposed. In the segmentation task for the ISIC2016 challenge skin lesion dataset, the model has further improved in segmentation accuracy compared with the original UNet++ model, which obtained 2.88%, 2.66%, 2.66%, 1.03%, 1.03%, and 1.66% in its six evaluation metrics of IoU, Recall, Precision, Accuracy, DICE coefficient and F1-score, respectively.