Leopoldo Corona

This is a protocol designed to optimize DE image subtraction for contrast-enhanced studies in rodents, employing iodine-based contrast medium (CM).

Our investigation used an Albira ARS commercial unit, not designed explicitly for quantitative CT tasks. DE subtraction was divided into stages that were independently analyzed: acquisition, volume reconstruction, image registration, and image weighting. The DE radiological techniques (low- and high- energy) had been previously optimized to enhance the visualization of iodine-based CM.

An independent reconstruction was needed to guarantee linearity between iodine intensity and its concentration for high energy acquisition; it also reduced structured noise occasionally produced by the microCT reconstruction software over uniform regions and improved bone visualization. Image registration was optimized, combining an affine transformation with a non-linear transformation determined with the Free-Form Deformation algorithm.

Two subtraction weight factors were identified: one that maximized the contrast-to-noise ratio (CNR) of iodine mixed with soft-tissue-equivalent resin and another that minimized CNR between bone-like rods and soft-tissue-equivalent material.

While working at Intelimetrica (a machine learning and data consulting startup), I was involved in the continuous improvement of a platform to manage mortgage portfolio quality.

This platform featured a house pricing model, with KPI of less than 5% of error, and a similar geographically-close houses finder model to avoid fraud in house appraisals. The similar houses model featured a similarity score as a secondary indicator of the quality of the appraisal.

I co-developed both machine learning models and contributed to the operationalization working close with the engineering team.

After starting my job early at Kueski (high-growing fintech startup financed by Silicon Valley VCs), we used a manual process of identity verification consisting of comparing the ID card face against a selfie. This process started to be unscalable and very prone to errors.

I proposed and developed an application that extracted the faces and inputted them to a model that returned the probability of being the same person to verify the loan applicant's identity automatically. This model was developed and trained from scratch using proprietary data, which had a state-of-the-art performance.