Undergraduate thesis · Systems EngineeringUniversidad Nacional José María Arguedas · 2023
Effectiveness of a Deep Learning classification model for COVID-19 detection
My thesis for the Systems Engineer degree. A modified DenseNet201 that detects COVID-19 in chest X-rays at 98% effectiveness, using 18.1M parameters and 3.2 GFLOPs.
COVID-19 spreads fast and is extremely contagious, which makes early diagnosis critical. At the time there was no effective tool to support detection alongside the standard tests — and that gap is what this thesis set out to address.
The work evaluates how effective and how efficient a deep-learning classifier can be at identifying COVID-19 from chest X-ray images. It uses three public datasets — COVID Chest X-ray, COVID Chest X-ray Dataset and the COVID-19 Radiography Database — plus a dataset of my own, kept aside for validation.
Several pretrained models from the DenseNet family were benchmarked, and from there I proposed a modification of DenseNet201: changes in the first convolution phase and in the transition layers, with transfer learning from ImageNet. The proposed model reaches 98% effectiveness with 18.1M parameters and 3.2 GFLOPs — better than the state-of-the-art proposals it was compared against, on both counts at once. That combination is the actual result. The model is accurate, and it is cheap enough to run somewhere with no GPU budget.
Presented in Andahuaylas in 2023 for the professional degree of Systems Engineer, advised by MSc. Herwin Alayn Huillcen Baca with MSc. Flor de Luz Palomino Valdivia as co-advisor. Open access in the UNAJMA institutional repository and indexed in ALICIA (CONCYTEC).