A Spatial-Temporal Multitask Deep-Learning Pipeline to Predict CT Perfusion Parameters

The Hong Kong University of Science and Technology
Department of Computer Science and Engineering


MPhil Thesis Defence


Title: "A Spatial-Temporal Multitask Deep-Learning Pipeline to Predict CT 
Perfusion Parameters"

By

Mr. Bohuai WU


Abstract:

Computed Tomography Perfusion (CTP) is a widely deployed imaging technique to 
assess cerebral perfusion in a wide range of central nervous system (CNS) 
diseases. The perfusion parameters CBV, CBF, TMAX and MTT are of the most 
clinical interest, traditionally derived from the CTP dataset by calculation 
through mathematical modeling. Despite its high accuracy, such calculation 
approach is prone to failure due to the presence of corrupted frames caused by 
artifacts (e.g., sporadic image noise or patient motion), and requires a high 
number of frames in a scan (and hence radiation exposure).

To overcome that, we investigate, for the first time, a multitask deep-learning 
approach to predict the perfusion parameters under the general realistic case 
of possibly non- uniform framing interval. We formulate a multitask learning 
problem to predict multiple perfusion parameters from CTP data simultaneously. 
Then, we propose STM DLP, a novel Spatial-Temporal Multitask Deep Learning 
Pipeline trained by the accurate calculation results as ground truth. STM-DLP 
consists of an Impulse Response Feature Encoder (IRFE) in the form of a spatial 
encoder followed by a temporal encoder, and a Multi-Parameter Predictor (MPP) 
which computes and outputs all the perfusion parameters in parallel. Extensive 
experiments on real CTP dataset demonstrate that STM-DLP predicts accurately, 
is robust against artifact failure, and requires much fewer frames (40% 
reduction).


Date:                   Wednesday, 22 May 2024

Time:                   3:30pm - 5:30pm

Venue:                  Room 4472
                        Lifts 25/26

Chairman:               Prof. Pedro SANDER

Committee Members:      Prof. Gary CHAN (Supervisor)
                        Prof. Raymond WONG