- This event has passed.
M.Tech Research: Thesis Defense: ONLINE: CDS: 28 October 2021″Novel Deep Learning Methods for Improving Low-Dose Computed Tomography Perfusion Imaging of Brain”
28 Oct @ 4:30 PM -- 5:30 PM
DEPARTMENT OF COMPUTATIONAL AND DATA SCIENCES
M.Tech Research Thesis Defense (Online)
Speaker : Mr. ARINDAM DUTTA
S.R. Number : 06-18-00-10-22-19-1-16729
Title : “Novel Deep Learning Methods for Improving Low-Dose Computed Tomography Perfusion Imaging of Brain”
Date & Time : 28th October 2021 (Thursday),04:30 PM
Venue : Online
Computed Tomography (CT) Perfusion imaging is a non-invasive medical imaging modality that has also established itself as a fast and economical imaging modality for diagnosing cerebrovascular diseases such as acute ischemia, subarachnoid hemorrhage, and vasospasm. Current CT perfusion imaging being dynamic in nature, requires three-dimensional data acquisition at multiple time points (temporal), resulting in a high dose for the patient under investigation. Low-dose CT perfusion (CTP) imaging suffers from low-quality perfusion maps as the noise in CTP data is spectral in nature. The thesis attempts to develop data-driven based deep learning algorithms to obtain improved perfusion maps directly from low-dose CT Perfusion data.
The inverse problem of obtaining high-quality perfusion maps from low-dose CT Perfusion data is a well-known ill-posed inverse problem. The present state-of-the-art techniques are computationally expensive and necessitate explicit information about the Arterial Input Function (AIF). To combat the same, we propose a novel deep learning-based end-to-end framework to produce high-quality Cerebral Blood Flow (CBF) maps from low-dose raw CTP data. The proposed models can perform the deconvolution without explicit information of the Arterial Input Function (AIF) and are not susceptible to varying levels of noise. Detailed experimentation and their results validated the superiority of the proposed deep learning framework over the existing state-of-the-art algorithms.
In the next part of the thesis, extension of this work was attempted by proposing a network which can handle variable number of time points. The novel hybrid network that we propose combines the benefits of three-dimensional (3D) and two-dimensional (2D) convolutions to handle variable number of time/temporal points was developed. It also performs deconvolution without explicit information of the Arterial Input Function (AIF). This is the first network that can handle variable time point dynamic 2D data and thus appeals to much wider use-case scenarios. Also, it can be extended to analogous modalities like Dynamic Contrast Enhanced-Magnetic Resonance Imaging (DCE-MRI).
The proposed methods are fully data-driven and are aimed at working with less training data, thus having a good appeal for clinical settings. They are single-step procedures with minimal preprocessing steps and provide fast processing without compromising the quality of the perfusion maps for low-dose CT perfusion imaging. Integrating these methods with the post-processing software platforms will enable the availability of high-quality perfusion maps especially for time critical operations like ischemic stroke imaging.