Can Image Enhancement be Beneficial to Find Smoke Images in Laparoscopic Surgery?

Congcong Wang, Vivek Sharma, Yu Fan, Faouzi Alaya Cheikh, Azeddine Beghdadi, Ole Jacob Elle, Rainer Stiefelhagen

Abstract

In laparoscopic surgery, surgical smoke often obscures the field of view, hindering both human surgeons and automated analysis tools. While desmoking algorithms can mitigate this, they are computationally intensive and require an accurate "trigger" mechanism. This paper explores whether targeted image enhancement can improve the performance of binary smoke classifiers. By utilizing a Weighted Least Squares (WLS) optimization framework to isolate and boost textural detail layers, the authors demonstrate a 4% improvement in smoke detection accuracy using a Support Vector Machine (SVM) on the Cholec80 dataset.

Introduction

Laparoscopic surgery relies on a clear camera feed. However, procedures like laser ablation generate smoke that creates glare and hides critical structures. To manage this efficiently, automated systems need a reliable "smoke detector"—a binary classifier that identifies smoky frames to trigger downstream processing. Current methods often struggle to distinguish subtle smoke textures from complex biological backgrounds. This research investigates if pre-processing images to make the "signature" of smoke more prominent can enhance machine learning performance.

Methods

The researchers developed an enhancement pipeline focused on feature discrimination rather than visual aesthetics.

The proposed WLS-based enhancement pipeline showing the multi-step fusion of detail layers.

The process involves three primary steps:

  1. Color Space Transformation: Images are converted from RGB to YCbCr to isolate the luminance (Y) channel, which contains the bulk of structural and textural information.
  2. WLS Decomposition: A Weighted Least Squares (WLS) filter is applied. This edge-preserving framework decomposes the image into "base" (coarse) and "detail" (fine) layers without blurring important boundaries.
  3. Multi-Scale Fusion: A two-level decomposition is performed using different smoothing parameters ($\lambda$). These layers are fused using strategies such as FC_AVG (averaging details) and FC_MAX (taking the maximum value to boost edges).

Following enhancement, 40-dimensional feature vectors based on Gradient Magnitude (GM) and Laplacian of Gaussian (LoG) are extracted and fed into a linear Support Vector Machine (SVM) for the final classification decision.

Datasets

The study utilized the Cholec80 dataset, a benchmark for gallbladder removal surgeries.

  • Training: 4,381 images from 3 videos.
  • Testing: 10,653 images from 9 videos.
  • Preprocessing: Images were downsampled to 427 × 240px to ensure computational efficiency during the enhancement phase.

Key Results and Insights

The proposed FC_AVG method emerged as the top performer, achieving an accuracy and F1-score of 0.64, representing a 4% improvement over the baseline RGB performance (0.60). It also outperformed specialized saturation-based methods like Saturation Analysis (SAN) by 1-5%.

Visual comparison of enhancement methods, highlighting how the proposed FC_AVG and FC_MAX approaches emphasize textural details.

A critical insight from this work is that images optimized for human visualization are not necessarily optimal for machine classifiers. The authors intentionally boosted high-frequency details that might appear "harsh" to a human eye but provide the SVM with essential textural cues needed to identify smoke. The WLS framework proved superior to standard Bilateral (BF) or Guided (GF) filters because it better preserves sharp edges while highlighting the hazy textures where smoke is most apparent.

Conclusion and Future Work

The study demonstrates that targeted image enhancement can significantly benefit automated surgical tools. The enhancement process takes approximately 0.32 seconds per image, making it a strong candidate for near real-time applications. Future work will focus on scaling the testing to the full 80-video Cholec80 dataset and integrating these enhanced images into deep learning architectures like Convolutional Neural Networks (CNNs) to further push the boundaries of detection accuracy.

Summary updated

BibTeX
@inproceedings{wang2018can,
  author = {Congcong Wang and Vivek Sharma and Yu Fan and Faouzi Alaya Cheikh and Azeddine Beghdadi and Ole Jacob Elle and Rainer Stiefelhagen},
  title = {{Can Image Enhancement be Beneficial to Find Smoke Images in Laparoscopic Surgery?}},
  booktitle = {CIC},
  year = {2018},
  doi = {10.2352/issn.2169-2629.2018.26.163},
}
Figure