Detect the "Undetectable"

Intro

Detect the “undetectable”; bold statement, and may even sound self-contradictory. In this context, I refer to “undetectable” as a section of a signal deemed undetectable via state-of-the-art tools/knowledge: It may be a tiny tumor only visible in a handful of pixels of a multimillion voxel 3D medical image or a mysterious pattern in financial data indicating that something interesting is about to unfold...

I am Engin Dikici, a scientist/inventor with educational & work experience applying AI and signal processing in various domains (e.g., medical, video, and financial data). I interpret and implement science as an art form that requires uninterrupted and complete dedication backed by a solid theoretical background, possibly the only way to detect the "undetectable" if it will ever be so.

Selected Projects

Brain cancer tumor detection

Detection of a small cancer tumor in the brain is like looking for a needle in a haystack: The 3D Computer Tomography (CT) data used by doctors to locate these malicious formations consist of tens of high-resolution images where a tumor could be just 1-2 pixels wide! The research focused on automated detection using an in-house Convolutional Neural Network developed for this challenging aim.
Dissemination(s): Paper and Patent

Predicting AI model generalizability

The generalizability of an AI system quantifies its performance for unseen data. The topic heavily attracts AI researchers' attention nowadays, as it is currently in its infancy stage without a clear understanding/solution. In this project, (1) a new loss function (i.e., sampled Frechet loss) regulating a model's latent space to a specific form and (2) a model generalization prediction system operating on this particular space are proposed.
Dissemination(s): Paper

Noisy-student-based Semi-supervised learning

Unlike many visual databases, it is not likely to have a labeled medical imaging dataset consisting of millions of samples; however, the researchers still aim to use SOTA (and highly parametric) DNN architectures in this domain. Yet, medical institutions hold high amounts of "unlabeled" data, commonly unutilized: This project aimed to exploit these untapped data reserves for developing AI models via a novel sensitivity-targetted Noisy-Student strategy.
Dissemination(s): Paper

Synthetic Volumetric Cancer Data Generation

Medical imaging data could be scarce in some scenarios (e.g., data collected for rare diseases), limiting data-driven imaging research. While being visually pleasing, the data generated from GAN algorithms commonly (1) fail with smaller training datasets and (2) could not be used for replacing original images in a research project. These limitations might be overcome using a constrained GAN ensemble formulation introduced in this research.
Dissemination(s): Paper and Patent

Augmented Deep Neural Networks

In addition to an AI model (e.g., DNN), most modern AI pipelines deploy various pre and post-processing stages that utilize the CPU. These might cause significant slowdowns in processing and a heavy workload on limited computational resources. The augmented network approach aims to move this workload to the GPU while improving the AI model's accuracy due to the combined training of all components.
Dissemination(s): Paper

Biomechanically Constrained Surface Tracking

The detection and tracking of organ borders in Ultrasound are challenging (due to the noisy nature of the modality). Integration of the biomechanical constraints could improve the performance of a tracking solution: The research focused on the (1) finite element analysis (FEA) of the Doo-Sabin subdivision surface models and (2) integration of this FEA into Kalman filter-based organ tracking framework.
Dissemination(s): Paper1 and Paper2

Ensemble Methods in Detecting Organ Borders

An ensemble of detectors is better than a single detector if and only if the statistical properties of each component are known. The project investigated and proposed multiple pathways to benefit from advanced ensemble strategies in detecting organ walls (i.e., James-Stein, Empirical Bayes, BLUE).
Dissemination(s): Paper1, Paper2 and Paper3

Automated vessel segmentation in 3D

The planning for the placement of stents and by-pass graphs may greatly benefit from the automated delineation of the vessel boundaries. Furthermore, the quantification of dimensions of an aneurysm may be critical in determining the time point for intervention. The project aimed to introduce graph cuts-based automated vessel segmentation algorithm for Computed Tomography.
Dissemination(s): Paper and Patent

Tracking and Analysis of Heart in MRI

The detection and tracking of the heart muscle (specifically the pumping chamber; myocardium) in MRI enables the computation of valuable medical information (e.g., ejection fraction). Furthermore, the damage to the heart tissue could be visualized via the delayed enhancement strategy: The project aimed to (1) track the 2D+T MRI and (2) perform the classification of infracted tissue via Machine Learning.
Dissemination(s): Paper and Patent

Elements

i = 0;

while (!deck.isInOrder()) {
    print 'Iteration ' + i;
    deck.shuffle();
    i++;
}

print 'It took ' + i + ' iterations to sort the deck.';