Theses and Dissertations

Permanent URI for this collectionhttp://ir.daiict.ac.in/handle/123456789/1

Browse

Search Results

Now showing 1 - 2 of 2
  • ItemOpen Access
    Tissue-Specific Analysis of Super Resolution Methods for Medical Images
    (Dhirubhai Ambani Institute of Information and Communication Technology, 2023) Doshi, Nisarg; Gohel, Bakul
    Image super-resolution (SR) techniques are widely used in various domains toenhance the resolution of low-resolution images, producing visually appealinghigh-resolution versions. However, regarding medical images, SR methods mustproduce precise results. Therefore, a thorough evaluation of the performance ofdifferent SR methods on various tissues is essential to determine their suitability.In particular, evaluating SR methods on region-specific organs, such as thelung, liver, and kidney in CT scans and brain in MRI scans, is essential. Whenthese organs are individually enhanced using Bi-cubic interpolation and Modified-ESPCN methods, along with standard evaluation metrics like Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM), it is observed that SRmethods exhibit inferior performance on most individual regions of interest comparedto the entire image. This difference in performance can lead to misleadinglyhigh results when evaluated over the entire image, which includes irrelevant nontissueregions.We propose using a tissue-specific model incorporating a region-based lossfunction to overcome this limitation. This approach allows for a more accurateand informative evaluation of SR methods in the context of tissue-specific performanceanalysis for CT images.
  • ItemOpen Access
    Super resolution of Covid-19 CT-Scan Images
    (Dhirubhai Ambani Institute of Information and Communication Technology, 2022) Patel, Vaidik Gautam; Gohel, Bakul
    Acquisition of high quality CT images is difficult, because it requires exposing patients to high doses of radiation. Super resolution algorithms can help in over coming this problem and obtain higher spatial resolution in CT images. Much deep learning based architecture have been proposed in the literature to overcome this problem. We perform the task of super resolution on a U-Net and study the effects of 2 preprocessing methods which are scaling and zscore. The evaluation strategy for the super resolution of CT images in the literature uses the Peak Signal to Noise Ratio (PSNR) and Structural Similarity (SSIM), however the results are published for the entire image. This is not a good practice for the evaluation of SR, we propose a novel region based similarity measurement practice and a lung specific or region of interest based similarity measurement. We further bifurcate the SSIM metric into it�s 3 component, i.e. luminance, contrast and structure, and study the impact of super resolution on each of these components.