CoSev: Data-Driven Optimizations for COVID-19 Severity Assessment in Low-Sample Regimes.

Area of Science:

• Computer Vision
• Medical Imaging Analysis
• Deep Learning

Background:

• COVID-19 diagnosis is well-studied, but severity assessment remains challenging.
• Limitations include small datasets, high-dimensional images, and GPU compute constraints.
• Accurate COVID-19 severity analysis is crucial for effective patient management.

Purpose of the Study:

• To develop an effective deep learning methodology for COVID-19 severity classification using 3D CT scans.
• To address challenges of high-dimensional data and limited sample sizes in medical imaging.
• To improve upon existing methods for COVID-19 severity assessment.

Main Methods:

• An iterative transfer learning approach was applied to 3D CT scans.
• Custom shallow convolutional neural network (CNN) architectures were designed and optimized.
• Models were trained progressively on simplified classification tasks (2-class to 5-class).
• Hyperparameters including learning rates, normalization, and dropout were systematically tuned.

Main Results:

• The developed model, CoSev, achieved an 81.57% classification accuracy on the MosMed Dataset.
• Performance improved significantly from initial accuracies below 60%.
• The methodology demonstrated state-of-the-art performance with simpler setup procedures.

Conclusions:

• Iterative transfer learning is effective for COVID-19 severity analysis on challenging 3D CT datasets.
• The approach advances computer vision for high-dimension, low-sample medical data.
• This methodology has potential applications in general disease detection and clinical practice.