Research

Computational Vision & Machine Intelligence
Computational Imaging Lab
Deep Learning Reading Group (DLRG)
Funded projects

My  work is mainly concerned with developing novel algorithms to understand how tumors affect tissue texture images. Different medical imaging modalities tend to characterize tissue with diverse types of textures, and there is no single approach to analyse all textures. A wise approach would be to investigate heterogeneity in tissue texture for signs of abnormality. A number of novel texture analysis techniques were developed to detect and differentiate between tumors subtypes early, and further assess their progression or regression.

The first texture is represented by images acquired noninvasively giving a fine texture structure, where lung tumor contrast enhanced images were acquired via computed tomography (CT) modality. Lung tumor staging predication accuracy was improved from conventional CT alone i.e. without using a PET scan, through identifying malignant aggressiveness of lung tumors by examining vascularized tumor regions that exhibit strong fractal characteristics. (more info)

Tumor texture analysis - Fractal parametric images
CT image and corresponding fractal dimension parametric images using the differential box-counting and fractal Brownian motion algorithms.

The quality of the extracted texture feature is substantial for an accurate diagnosis. Therefore the impact of noise extracted via different statistical, model and filtered-based texture analysis algorithms was assessed on non-contrast enhanced CT images. Experiments were performed on lung tumor regions of interest, and the performance of the texture analysis method was compared without and under noise presence. (more info)

Tumor texture analysis - CT noise
[left] Original CT image and corresponding noise (indicated by yellow arrow), [right] horizontal profile along the middle region of interest for reconstructed clean and noisy CT image versions.

In another work, a different type of texture was dealt with for the purpose of developing an automated system for improving histopathological meningioma brain tumors discrimination acquired invasively via digital microscopy modality. Histopathological texture usually has a macro or coarse structure where the cell nuclei color, shape and orientation defines the general texture structure and play an important role in the feature extraction process (more info) and for automated classification (more info, video).

Tumor texture analysis - histology texture analysis
Meningioma histopathological images and corresponding fractal dimension parametric images.

My work was also concerned with a third type of texture represented in the problem of ultrasound tissue characterization. Tumor spatial and contrast resolution in ultrasound images is low as compared to other modalities, therefore a new approach for assessing subtle heterogeneities within a given mass in ultrasound image texture was proposed. Volumetric Nakagami-based shape and scale parameters were estimated from ultrasound RF, and a multi-resolution Daubechies wavelet packet transform was performed adaptively. Finally, local multi-scale textural fractal descriptors were extracted from volumetric patches. Results show improved prediction of therapy response and tumor characterization. (more info)

Tumor texture analysis - fractal map - Nakagami shape

Tumor texture analysis - fractal map - Nakagami scale
Multiresolution fractal slice maps referring to, respectively, Nakagami shape and scale parametric voxels of radio-frequency ultrasound images, where blue regions indicate tumor response to chemotherapy treatment.

Also fractal analysis of the echo signal was found to give additional information about the heterogeneity of the underlying liver tissue structure. Regions within the liver tumor tissue which responded to chemotherapy treatment were shown to exhibit different statistical properties to that of the non-respondent counterpart. (more info)

Tissue_heterogeneity
Ultrasound radio-frequency tissue characterization via B-mode versus parametric imaging.

In the realm of myocardial segmentation, a novel algorithm was proposed for correct left ventricle (LV) segmentation in 3D echocardiographic images sequences based on fractional Brownian motion. Dealing with the LV segmentation problem from a spatio-temporal perspective can give further information on the shape boundaries. (more info)

Left Ventricle segmentation
Cardiac ultrasound image segmentation using fractional Brownian motion.