ENSEMBLE APPROACH FOR IMPROVING KIDNEY TUMORS SEGMENTATION PERFORMANCE ON CT IMAGES USING DEEP LEARNING MODELS

Abstract

Clinical picture examinations are conducted with the help of image segmentation, which allows the computerized picture split into a set of pixels. In segmentation, the aim is to enhance and modify the delineation of a picture so that it becomes more distinguished and easier to investigate. Kidney growths address a sort of malignancy that individuals of old age are bound to create. In this respect, Deep Learning (DL) models are becoming increasingly appealing. Creating models for kidney tumor segmentation assist doctors/radiologists in recognizing cancers with effective division as an integral step. A comparison of the segmentation approaches using Attention U-Net, Feature Pyramid Network (FPN) and LinkNet Models is presented in this paper to develop the ideal prediction model on Computed Tomography (CT) images. Various encoders are used in all three architectures to build different predictor models. Ensemble approach using Attention-U-Net architecture outperforms compared to FPN and LinkNet architectures with IoU scores 95.66 %( kidney) and 93.86% (tumor).