Impact of melanocytic lesion image databases on the pre-training of segmentation tasks using the UNET architecture

Image segmentation is a fundamental task for automatic disease diagnostics, and UNET is a widely used architecture for medical image segmentation. This is because this architecture is relatively simple to implement and obtains robust results. Currently there are several databases of melanocytic lesions, so determining which databases and the amount of data needed to train on such an architecture is a challenge. Another drawback is often to generate a proper database, since the ideal segmentation needed for training requires a large effort on the part of medical professionals. In this paper, we analyze the impact of melanocytic lesion databases on pretraining using the UNET architecture and its impact on segmentation results. ResNet with pre-trained weights with Imagenet is used. Three different databases were used in the experiments. Experimental results show that models pre-trained with images of melanocytic lesions perform better than those models that do not use pre-training for segmentation. In turn, adding more models pre-trained with melanocytic lesion database gives better results in most cases. Finally, for most of the tests performed it can be observed that using Imagenet only in pre-training obtains worse results than combining pretraining with other melanocytic lesion databases.