2024

Bachelor’s thesis

on the topic:

“Improvement of the signal-to-noise ratio of optical coherence tomography images using artificial intelligence methods”

Leonhard Knipfelberg

Sharper images drive better decisions – in medicine, biology, and beyond. Optical coherence tomography (OCT) can reveal micrometer-scale structure, yet characteristic noise (especially speckle) often blurs detail and biases measurements.

This thesis benchmarks four denoising methods on a self-acquired OCT dataset of plants to guide OpenLabKI, an open-source toolkit for OCT image processing. The methods span classical and learning-based approaches: median filtering, BM3D, Noise2Noise, and Noise2Void. The central question is practical: which method offers the best trade-off between noise suppression, structure preservation, and computational cost for routine OCT workflows?

Performance is assessed using complementary criteria that are interpretable across disciplines: PSNR (signal fidelity), SSIM (perceived structural similarity), RMSE (error magnitude), and per-image processing time. These metrics capture both quantitative accuracy and user-facing speed. Implementations are provided in Python with standardized preprocessing and a reproducible evaluation pipeline, enabling like-for-like comparison.

OCT is widely used in ophthalmology and increasingly in plant and material sciences. Regardless of domain, high noise undermines contrast, reduces effective signal-to-noise ratio, and can degrade downstream analysis or diagnosis. By systematically comparing denoising strategies on real OCT volumes (here, plant tissue as a patient-free proxy for complex microstructure), this work aims to (i) quantify how much each method improves image quality without erasing fine features, and (ii) recommend a default algorithm for OpenLabKI that balances quality with inference time on typical hardware.

Contributions are threefold: a clear, metric-driven benchmark of four representative denoisers; practical integration guidance for OpenLabKI; and open, modular code that can be extended to new datasets or modalities. The outcome is a reasoned recommendation – grounded in measurable gains and runtime constraints – for an OCT denoising module that non-specialists can adopt with confidence.