Hybrid Approach for Multiple Organ segmentation and Tumor detection

Abstract

Accurate and fast segmentation of medical images is one of the fundamental newlineprocesses used for disease diagnosis, disease progression tracking, selecting suitable newlinesurgical procedures, and radiation therapy. Manual segmentation is most commonly used newlineand trusted method in medical practice. However large number of images generated in newlineclinical routine makes it difficult for manual segmentation. In addition, manual newlinesegmentation is time consuming, subjective and depends on the level of individual s newlineexperience. Therefore, automated, accurate and reliable segmentation methods are newlinerequired. However, heterogeneity of tissue cells, lack of clear boundary information, large newlinespatial and structural variability and scarcity of annotated training data makes automatic newlinesegmentation a very challenging task. This research work investigates the machine learning newlinemethods for two key challenges in medical image analysis: The first one is segmentation newlineof organs and tumors from biomedical images. The second one is organs localization for newlineautomating the shimming localization process of MRI. newlineThe first main contribution of the thesis is, a series of novel approaches using newlineadvanced concepts deep learning are introduced for medical image segmentation. newlineCombination of U-Net with Residual blocks, Inception modules and Dense-Net are newlineinvestigated for semantic segmentation of organs and tumors from MRI and CT images. newlineApplications are chosen Kidney, Prostate, Liver, Brain and Bone scans and efficacy of the newlineproposed algorithms are demonstrated on several publicly available data sets. Dice newlineSimilarity Coefficient, Jaccard index and hausdroff distance metrics are used to evaluate newlinethe results of proposed models. The preliminary results show that deep learning-based newlinetechnique outperforms all existing traditional segmentation algorithms. newlineThe second main contribution of the thesis is organs localization on MRI images to newlineperform automated localization of shimming. Artifacts due to inhomogeneous magnetic newlinefields is very common in MR images,