Jiahang Su

In the last few years, the blooming of learning based methods has brought reincarnation to the medical image analysis field. Classical methods were rapidly replaced by many kinds of learning based methods. This thesis has explored biomarker extraction methods and applications in ischemic stroke. Such biomarkers may be used in treatment decision making. The first chapters of this thesis focuses on collateral scoring, including model design, the assessment of model performance and the clinical use of the automatic collateral scoring method. In of the last chapters this thesis, a general post-processing method that improves binary segmentation results for the tubular structure was discussed, and a method that extracts the anterior trees from CTA images was developed and thoroughly assessed.

In Chapter 2, we designed a method for collateral scoring that follows the human visual collateral scoring approach. The proposed method consists of three steps. First, the brain region is defined using atlas based registration. Next, a 3D U-net for vessel segmentation is applied. Finally, the collateral score is determined using the median ratio of vessel features on the occluded side versus the contralateral side. The obtained collateral score is a floating point score. The floating point scores were converted into categorical scales using a simple regression model or random forest classifier. The performance of the collateral scoring method was assessed in a randomly sampled subset of a large multicenter registry dataset (MR CLEAN Registry). The subset consist of 270 subjects and the collateral reference standard was a consensus score obtained from three experienced radiologists. The performance of the proposed method is comparable to an experienced radiologist. In addition, a manually labeled cerebral vessel centerline dataset was created for training the vessel segmentation model.

In Chapter 3, we designed an end-to-end approach for modeling collateral score biomarker. The proposed method treats collateral scoring as a classification problem using a Siamese model to extract features and to compare the occluded hemisphere and healthy hemisphere. For ease of direct comparison, the occluded hemisphere and healthy hemisphere were aligned using atlas based registration. The output of this method was a value on a four-point scale. This method was evaluated using data from a randomized control trial, the MR CLEAN trial dataset. The performance of this end-to-end approach is similar to the method in Chapter 2.

In Chapter 4, we validate the performance of the proposed method in Chapter 2 with 29 raters (either radiologist or radiology resident). There is no statistically significant difference between the accuracy of human raters and the accuracy of the automatic collateral scoring method.

In Chapter 5, we use the floating point automatic collateral scoring method from Chapter 2 to investigate the optimal contrast acquisition time point for CTA images by computing the CS over all timepoints of CTP and mCTA images. This information is used to temporarily align CTP and mCTA images of the same subjects. The study shows that collateral scores greatly depend on the timing of the acquisition, that mCTA images can be accurately timed, and that, when using baseline parameters as a metric, a CTP image does not have added value for collateral scoring; a good-timed CTA image is sufficient.

In Chapter 6, a general method that is intended to improve the result of a tubular structure segmentation (e.g. a lung airway segmentation result and brain vessel segmentation result) is presented. The method learns to improve the initial segmentation results by training with data from a GAN based synthetic error generator. This method can act as an add-on to the initial segmentation approach. The connectivity and completeness of the tubular structure segmentation result improve statistically significantly when applying this method. The manual labelled dataset created Chapter 2 was used in this work.

In Chapter 7, a method for extracting the cerebral anterior vessel tree was presented. The method consists of three parts, a policy gradient based deep reinforcement learning (DRL) tracker, a CNN based bifurcation detection, and a classical breadth-first tree formation method. In the DRL tracking approach, we adopted a proximal policy optimization deep reinforcement learning method in an online fashion. A curve-to-curve distance metric reward function was proposed. We also investigated the network architecture configuration and the training scheme. In addition, we proposed a metric to better monitor the training process of the DRL tracker rather than relying on the accumulated reward. The bifurcation detection used a similar architecture but with switchable normalization after each convolutional layer. At last, a breath-first tree formation method with a tracking ensemble method was proposed to extract the anterior tree. The application of anterior tree extraction is novel in the field and the performance of DRL tracker achieves state of art performance. Moreover, a dataset of manually labeled cerebral anterior trees over 125 subjects (randomly selected from MR CLEAN Registry) was created for assessing and training the proposed method.