Home > Events > Oral Candidacy - Ali Shahbazi

Oral Candidacy - Ali Shahbazi

Start: 1/29/2018 at 1:00PM
End: 1/29/2018 at 4:00PM
Location: 315 Stinson Remick
Add to calendar:
iCal vCal

Ali Shahbazi

Oral Candidacy

January 29, 2018        1:00 pm         315 Stinson

Adviser:  Dr. Walter Scheirer

Committee Members

Dr. Kevin Bowyer        Dr. Adam Czajka        Dr. Bobby Kasthuri


Computer Vision-based Approaches to Neural Circuit Tracing at Scale 


New developments in imaging techniques such as electron microscopy imaging, micro-CT X-ray and the recently proposed spectral confocal reflective imaging approach facilitate the investigation of fusion between brain structural map-ping and its functional or behavioral activities. These methods utilize features such as synapse detection, synaptic conjunction mapping and micro-structure nanoscale imaging of cells and vasculature. Manual annotation techniques were initially used to analyze Nano-scale neuro-images. However, these methods are considerably slow and expensive when applied to image stacks on the scale of gigabytes.

Novel computer vision approaches such as supervised machine learning and deep learning were introduced to overcome the aforementioned shortcomings. These techniques required sizable manually expert annotated stacks (blocks of images) for training and evaluation purposes. These stacks should run through supercomputers or large grids powered by CPU/GPU computing infrastructure to reach a desirable result. To study the feasibility of computer vision in this domain, we proposed an interdisciplinary project between computer vision and neuroscience to automatically map circuit motifs and microstructure of the brain, and to accelerate the reconstruction of a more prominent volume of brain structures by making a faster tool based on a computer vision approach. Imaging is a dominant strategy for data collection in neuroscience, yielding 3D stacks of images that can scale to petabytes of data for a single experiment, depending on the modality. Machine learning-based algorithms from the realm of computer vision can serve as a pair of virtual eyes that tirelessly processes these images to construct flawless and realistic circuits automatically. However, in practice, such algorithms are more often than not too error-prone and computationally expensive to be immediately useful. Therefore we introduce new fast, flexible learning-free automated method for sparse segmentation and2D/3D reconstruction of brain microstructure.

My proposed works will start with our flexible learning-free approach for segmenting and reconstruction of brain micro-structures. Different from prior supervised methods, our algorithm exploits cell-specific context clues and requires no extensive pretraining. Our approach works on various modalities and sample targets, including serial section electron microscopy (EM) of APEX2-positive processes, high-energy synchrotron X-ray micro tomography (μCT) of cortical volumes and invivo spectral confocal reflectance microscopy. Experiments on newly published and novel mouse datasets demonstrate high precision and recall for the proposed algorithm, as well as reconstructions of sufficient quality for further biological work. Compared to existing supervised and unsupervised methods, it is both significantly faster (up to several orders of magnitude) and comparable in segmentation and reconstruction performance. The proposed work will add new and novel features to the existing software to make a suitable package for neuroscientists to use for segmentation and re-construction. We also plan to make an extensive training set database for different modalities and resolutions by data augmentation over available annotated datasets. This package will be available as an online tool to help scientists in various fields reconstructs their own dataset.