AAU Student Projects is unavailable between June 15th 1.30pm and 17th 1.30pm due to planned system maintenance. The projects cannot be downloaded during this period.
AAU Student Projects - visit Aalborg University's student projects portal
A master's thesis from Aalborg University
Book cover


A Study of Reliability-Aware Super-Resolution in Robotic Laparoscopic Imaging

Author

Term

4. semester

Publication year

2026

Submitted on

Pages

85

Abstract

Laparoscopic cameras used in minimally invasive surgery often have limited spatial resolution because of size and hardware constraints. Super-resolution (SR) uses algorithms to reconstruct higher-detail images from lower-resolution inputs. However, simply increasing apparent resolution can invent details that are not real ("hallucinations"), so clinical use also requires showing how trustworthy the reconstruction is. This thesis explores a deep learning SR pipeline that aims to improve image quality while explicitly modeling uncertainty and evaluating reliability. A modular framework brings together: (1) representative SR models (EDSR and Swin2SR) for reconstruction, (2) uncertainty estimation, and (3) reliability assessment. Uncertainty is estimated with sensitivity-based maps obtained by perturbing the input or the inference process and observing how the output changes. Conformal prediction is used to build calibrated reliability masks that indicate where in the image the model’s output can be trusted. A conceptual, unified reliability score combines fidelity, perceptual quality, and uncertainty into a single measure. Experiments indicate that Swin2SR performs best among the tested models, that an inference-time transform perturbation captures uncertainty behavior well, and that reliability varies across the image, especially in complex regions. Overall, the thesis argues that medical-image SR should be treated as a reliability-aware imaging problem, not just an image enhancement task.

Mange laparoskopiske kameraer i minimal invasiv kirurgi har begrænset opløsning på grund af størrelse og hardware. Superopløsning (SR) bruger algoritmer til at genskabe mere detaljerede billeder ud fra lavere opløsning. Men at øge den tilsyneladende opløsning kan også opfinde detaljer, der ikke er reelle ("hallucinationer"), så klinisk anvendelse kræver metoder, der også viser, hvor troværdig rekonstruktionen er. Denne afhandling undersøger en dyb-læringsbaseret SR‑pipeline, der skal forbedre billedkvaliteten og samtidig eksplicit modellere usikkerhed og vurdere pålidelighed. Et modulært rammeværk samler: (1) repræsentative SR‑modeller (EDSR og Swin2SR) til rekonstruktion, (2) usikkerhedsestimering og (3) pålidelighedsevaluering. Usikkerhed estimeres med følsomhedsbaserede kort, der opnås ved at perturbere input eller selve inferensprocessen og observere, hvordan output ændrer sig. Konformal prædiktion bruges til at lave kalibrerede pålidelighedsmasker, der angiver, hvor i billedet modellens output kan stoles på. En konceptuel, samlet pålidelighedsscore kombinerer fidelitet, perceptuel kvalitet og usikkerhed i ét mål. Resultaterne viser, at Swin2SR opnår den bedste præstation blandt de testede modeller, at en inferens‑transform‑perturbation beskriver usikkerhedsadfærden godt, og at pålideligheden varierer rumligt, især i komplekse områder. Overordnet viser afhandlingen, at SR til medicinsk billeddannelse bør behandles som et pålidelighedsbevidst billedproblem og ikke kun som billedforbedring.

[This apstract has been rewritten with the help of AI based on the project's original abstract]