Transfer learning based framework for image segmentation using medical images and Tversky similarity

Abstract Medical image segmentation and further processing are extremely difficult due to the variations in images produced by various medical imaging techniques. In addition, characteristics like color, size, shape, and the separation of the foreground and background in an image pose additional challenges. Innovative ideas and efficient designs are needed to provide more accurate results. The size of the labeled dataset also significantly impacts how effectively the proposed deep learning architecture-based model works. To achieve better segmentation results using skin cancer datasets, this paper describes a method for implementing the deep learning framework using pre-trained models. This paper proposes the MU-Net, in which the encoder part of U-Net has been replaced with a MobileNetV2 model that has already been trained and uses a focal tversky loss function to control data and class imbalances and obtain high segmentation accuracy. This method is better than existing models as it has a less number of parameters (and thus requires less computation) and is less prone to overfitting, which is a menace for medical image datasets (as most of them are small). The efficiency of the proposed work has been analyzed by comparing its performance with existing models and evaluated based on performance metrics. It surpasses the competition with high accuracy (99%) and low loss (6%).The acclaimed deep learning architecture has improved performance and is more beneficial for segmentation and classification in other domains.