Welcome!

Gustavo Garzón is a PhD student of Computer Science at Universidad Industrial de Santander (Colombia). His research interests include: medical image processing, action recognition, pattern recognition, computer vision, motion analysis, image processing and machine learning.

He works as adjunct professor at Universidad Industrial de Santander on Automata (formal languages) and Artificial Intelligence courses.

Also, at Biomedical Imaging, Vision and Learning Laboratory (BIVL2ab) collective as Research Assistant and helps with student advising on computer vision related areas.

Current research

A synthesis of DWI sequences and ADC maps from NCCT onset sequences following ischemic stroke lesion-informed priors (2024)

Immediate profiling of ischemic stroke is crucial for providing endovascular or pharmacological treatment, and avoiding neurological repercussions. Current stroke care protocols use a non-contrast computed tomography (NCCT) to discard other stroke mimic conditions, but showing low differentiation between healthy and affected tissue. The protocols may include complementary findings from diffusion-weighted MRI (DWI/ADC) to improve analysis. However, DWI's low availability and patient conditions may limit its acquisition. This work proposes a generative representation capable of synthesizing ADC sequences from NCCT studies, forcing training to focus on stroke-affected regions while preserving textural characteristics of ischemic lesions. The approach synthesizes ADC images from a standard encoder-decoder net, updating the output from a complementary auxiliary encoder to learn ischemic stroke lesions. A retrospective study accumulated 196 NCCT and MRI sequences of patients diagnosed with ischemic stroke, acquired in a time interval less than 24 hours. A subset of 96 patients was considered for training and testing the proposed model, while the remaining 100 patients were used to conduct a perceptual study with five external radiologists. The proposed method synthesized ADC maps with an average reconstruction (SSIM/PSNR) score of 0.88/25.32 and ischemic lesion reconstruction performance of 0.67/20.02. Further, in a perception study, experts assessed the perceived realism of synthetic and real sequences. A Levene test showed a non-statistical difference between real and synthetic sets ($\rho>0.1$) from scores brought by radiologist observations. The proposed method can synthesize approximations of MRI sequences (DWI/ADC) from NCCT scans, potentially reducing time delay between image acquisitions and providing estimates of ischemic lesions observable only in MRI sequences.

Current research

A deep CT to MRI unpaired translation that preserve ischemic stroke lesions (2022)

Stroke is the second-leading cause of death world around. The immediate attention is key to patient prognosis. Ischemic stroke diagnosis typically involves neuroimaging studies (MRI and CT scans) and clinical protocols to characterize lesions and support decisions about treatment to be administered to the patient. Nowadays, multiparametric MRI images are the standard tool to visualize core and penumbra of ischemic stroke, supporting diagnosis and lesion prognosis. Specially, DWI modality (Diffusion Weighted Imaging) allows to quantify the cellular density of the tissue, and therefore allowing to quantify the lesion aggressiveness, and the recognition of micro-circulation properties. Nevertheless, MRI availability at hospitals is not widespread, and acquisition require special conditions requiring considerable time. Contrary, CT scans commonly have major availability but brain structures are poorly delineated, and even worse, ischemic lesions are only visible at advanced stages of the disease. This work introduces a deep generative strategy that allows ischemic stroke lesion translation over synthetic DWI-MRI images. This encoder-decoder architecture, include U-net modules, hierarchically organized, with inter-level connections that preserve brain structures, while codifying an embedding representation. Then a cyclic loss was here implemented to receive CT inputs and decode DWI-MRI images. To avoid mode collapse, this learning is inversely propagated, i.e., from synthetic DWI-MRI images to original CT-scans. Finally, an embedding projection is recovered to show a proper lesion-slice discrimination, regarding control studies. Clinical relevance- To recover synthetic DWI-MRI that preserved ischemic lesion using CT scans as an input and following an unpaired image translation setup.

Current research

Characterization of ischemic stroke lesion over complementary images modeled as deep generative representations (PhD. thesis)

Projects

Spatio-temporal patterns representation for action recognition (2019)

This work proposes an efficient gesture recognition descriptor that processes salient points represented from improved motion trajectories on each frame. A Poisson grid is adjusted and point-ocurrences are stored into a histogram. Histograms of Optical Flow (HOOF) and Motion Boundary (MBH) are then added for robustness, resulting in a joint spatio-temporal motion descriptor. The proposed motion descriptor shows to be competitive to represent gestures in videos, robust to local changes on movement and shows a low computational cost.