Burst Super-Resolution for Intraoperative Image Analysis Without High-Resolution Ground Truth

Abstract

Intraoperative imaging, despite advances in high-resolution sensors, often suffers from effective resolution loss due to motion blur, suboptimal illumination, acquisition distance, and real-time compression, reducing image clarity and limiting reliable computer-aided analysis during minimally invasive surgery (MIS). Super-resolution (SR) offers a potential solution, but most existing approaches focus on single-image restoration evaluated on synthetic degradations within the same domain, limiting applicability to real surgical scenarios. We introduce a pipeline for evaluating and adapting burst-based multi-image super-resolution (MISR), using short sequences of consecutive low-resolution frames, for intraoperative video enhancement under realistic clinical constraints. The study uses a private dataset of monochannel videos from 289 surgical procedures to test algorithms under real clinical conditions where high-resolution ground truth is unavailable. We examine four MISR architectures: a self-supervised model (DSA) and three supervised approaches, transformer-based (Burstormer, TR-MISR) and generative (ESRGAN), adapted for $2\times$ upscaling on monochannel video. Most models improve image quality with additional frames, though gains vary by architecture and metric. Taking the median across all five no-reference metrics, the best burst configurations yielded improvements of up to 14% (Burstormer), with more moderate gains for TR-MISR and ESRGAN (3–4%), while DSA showed no consistent benefit. Clinical experts preferred multi-frame reconstructions over single-image and original low-resolution images in 63% of cases, versus 32% and 5%. This work 1) demonstrates the feasibility of applying MISR to real surgical video, 2) presents a methodology for domain-adaptive evaluation without high-resolution references, and 3) offers insights into MISR behavior under clinical constraints, supporting robust super-resolution strategies.