Medical image segmentation using novel u-net variants

Abstract

In the realm of medical imaging, precise segmentation holds the key to unlocking invaluable insights for diagnosis and treatment. Leveraging the power of deep learning, this thesis explores innovative methodologies to enhance the accuracy and efficiency of medical image segmentation. Recently, deep learning models, particularly U-Net-like architectures, have obtained notable success in medical image segmentation. Ensemble learning is a powerful approach that helps leverage the performance of an overall model by incorporating the decisions of multiple models. We propose an ensemble learning model, called BLUE-Net, using a combination of three models, namely the basic U-Net, Link-Net, and FPN. We combine the decisions of these models with a fuzzy-ranking scheme by using a modified Blumberg function to decide whether a pixel in an image is a part of a foreground pixel of the segmentation mask. We also propose a novel model, called SEAU-Net, which employs attention to the lower skip connections only, effectively integrating SE attention blocks within the skip connections of the encoder backbone. Highlevel semantic features are ensured by these attention-aided feature-maps from the lower layers to the higher layers in the decoder that are effectively utilized for the reconstruction of segmentation maps. Recognizing key features from input images is a potent strategy to enhance overall model performance. One way to achieve this is by integrating decisions from multiple outputs. Hence, we introduce the W2CU-Net, which is a dual decoder network based on Cosine Similarity, Weibull function and Weierstrass’ Elliptic Curve Equation. It is designed for liver and liver tumor segmentation from CT scan images. The model employs an ensemble scheme, which combines the decisions of two decoders using a modifiedWeierstrass’ elliptic curve equation while selecting important features using Cosine Similarity and Weibull function, to determine whether a pixel in a CT scan image belongs to the foreground of the segmentation mask. Three models are evaluated across six datasets: 3DIRCADb, CHAOS, LiTS for liver and liver tumor segmentation, BUSI for breast cancer segmentation, OCDC for oral cancer segmentation, and ISIC 2018 for skin cancer segmentation.