Source-linked AI summary
Cross-dataset harmonization for robust endoscopic image analysis
Romil Imtiaz, Dimitris K. Iakovidis
TL;DR
Endoscopic models often underperform across devices, where sensors and illumination shift image appearance and color. The paper applies reference-based CIE-Lab harmonization to target images without retraining; it reports consistent cross-dataset gains, including a 30.7 increase in AUC for one evaluated setting.
Problem
Endoscopic models often underperform on images from devices different from their training systems, where device- and pipeline-dependent appearance shifts include changes in global illumination and color rendering.
Source: §1 · p. 1, §3.1 · p. 2Method
The method maps target images toward a source/reference dataset’s appearance using fixed CIE-Lab statistics and deterministic preprocessing, without retraining.
Source: §1 · p. 1, §3.1 · p. 2Results
Harmonization consistently improved cross-dataset generalization for detection and classification; on DS3, InceptionV3 AUC rose from 71.4% to 93.3% (+30.7).
Source: §4.1 · p. 4, §5 · p. 5Takeaways & Limitations
The same trained weights and inference settings can be used across datasets after offline harmonization, which adds no inference-time cost to the deployed model.
Source: §1 · p. 1, §3.2 · p. 3
Abstract
from arXiv · showhide
A significant problem in endoscopic image analysis is that the machine learning (ML) models used for this purpose usually underperform when applied on images acquired from endoscopes that are different from those used to acquire the images of their training set. The main difference of the images originating from different endoscopes is their color distributions, which depend both on the image sensors and the light sources used. Although previous studies have highlighted this challenge, to the best of our knowledge it has not been previously explicitly tackled. This study focuses on this problem and proposes very simple but impactful method. It implements a reference-based image harmonization that reduces global appearance differences between endoscopic datasets. Specifically, it extracts global color statistics from a chosen reference dataset in the CIE-Lab color space and applies a statistical channel-wise transformation to map each target image toward the appearance of the images of the reference dataset. The method is evaluated in the context of polyp detection in both flexible colonoscopy and capsule endoscopy datasets using a dataset-level cross validation protocol. The results indicate that the proposed harmonization consistently improves cross-dataset performance up to 30.7%, outperforming relevant baseline and state-of-the-art methods. The results indicate that a substantial part of the generalization gap is driven by low-level appearance variation that can be mitigated without retraining.
1. INTRODUCTION
Endoscopic models can lose performance across devices because sensors and illumination alter image appearance, especially color. The paper proposes a reference-based harmonization approach and frames it as an endoscopy-specific preprocessing contribution.
1. INTRODUCTION
- Cross-device shifts in global appearance, especially color distributions driven by sensors and illumination, are a major cause of endoscopic model degradation.Acquisition and processing differences may also create domain shift that could contribute to medical-image misinterpretation in practice.Source: §1 · p. 1
- Fig. 1 juxtaposes samples from five endoscopy datasets with harmonized versions, using sample (a) as the reference.Source: Figure 1 · p. 1
- The method maps target images toward a reference dataset’s appearance using CIE-Lab color statistics and a channel-wise statistical transformation.The fixed statistics enable preprocessing with the same trained weights and inference settings, without retraining.Source: §1 · p. 1
- The contribution is an endoscopy-specific preprocessing protocol; the authors report that no other state-of-the-art method explicitly tackles this problem.They distinguish it from training-heavy adaptation frameworks.Source: §1 · p. 1
2. RELATED WORK
Prior endoscopy work has enhanced individual images or addressed cross-device differences through normalization, color transfer, adaptation, and synthetic data. The paper distinguishes its approach as deterministic reference-statistics preprocessing without retraining or manual method choices.
2. RELATED WORK
- Acquisition-device heterogeneity complicates cross-dataset evaluation and pooling, while dataset metadata and usability also vary.The cited review reports inconsistent capture details; the related-work passage names capsule and flexible-colonoscopy datasets as representative resources.Source: §2 · pp. 1–2
- Per-image enhancement methods improve color, contrast, or tissue detail, but do not align tissue appearance across devices.Examples include fusion-based enhancement, TXI, color reproduction, and video color transformations.Source: §2 · p. 2
- Cross-device alignment methods include Z-score normalization, histogram matching, FDA, PCA/SVD color transfer, and palette-based transfer.These approaches align intensity distributions, low-frequency appearance, dominant color components, or color groups.Source: §2 · p. 2
- Other domain-shift approaches add adaptation mechanisms and training design or use generative models to expand training data.The passage cites domain-adaptive colonoscopy frameworks and synthetic-data augmentation.Source: §2 · p. 2
- Unlike those alternatives, the proposed method uses fixed CIE-Lab reference statistics for deterministic target-image transformation without retraining or manual parameter choices.FDA, PCA/SVD, and palette-based transfer require choices such as spectral region size, transformation strength, color space, or palette size.Source: §2 · p. 2
3. METHODOLOGY
The method harmonizes target datasets toward a source/reference dataset’s appearance while keeping the model pipeline fixed. It uses fixed CIE-Lab reference statistics and transforms each target image independently as offline preprocessing.
3. METHODOLOGY
- The dataset-level harmonization step compares original and transformed target inputs while leaving model architecture, training, and learned parameters unchanged.This design isolates input appearance as the changed factor in performance comparisons.Source: §3 · p. 2
3.1. Problem Setting and Harmonization Objective
- The deterministic transformation maps target dataset D_t into the appearance space of source/reference dataset D_s without retraining or parameter optimization.It can be applied offline as preprocessing, with the learning pipeline unchanged.Source: §3.1 · p. 2
3.2. Image Harmonization
- Reference channel means and standard deviations are computed once from all pixels in the training/reference dataset in CIE-Lab space.Each target image is converted to Lab, and its channel-wise statistics are mapped to the fixed reference statistics.Source: §3.2 · p. 2
- The channel transformation normalizes each target channel using its per-image statistics, restores it with reference statistics, clips Lab values, and converts the image back to RGB.A small ε=10^-6 is added to standard-deviation terms to avoid numerical instability and division by zero; the algorithm repeats the procedure across target images and saves them using the original naming convention.Source: §3.2 · pp. 2–3
- The controlled preprocessing keeps detector weights, inference settings, and evaluation pipelines fixed, and adds no inference-time cost to the deployed model.Performance changes are therefore attributed to input appearance rather than retraining or parameter tuning.Source: §3.2 · p. 3
4. RESULTS AND DISCUSSION
The evaluation compares harmonized and raw images across flexible-colonoscopy and capsule-endoscopy targets while keeping the learning pipeline fixed. Results show improved cross-dataset performance, with the largest reported AUC gain on DS3.
4. RESULTS AND DISCUSSION
- For DS2 and DS3, models trained on DS1 were tested on independent targets; the comparisons held weights, inference settings, decision threshold, and evaluation pipeline fixed.The experiments also tested DS4-to-DS5 transfer, with sensitivity, specificity, F1-score, and AUC evaluated.Source: §4 · p. 3
4.1. Flexible colonoscopy
- Figure 2 contrasts raw DS2/DS3 images and failed detections with harmonized images and successful detections, alongside DS1 reference images.Source: Figure 2 · p. 3, §4.1 · p. 4
- On DS2, improvements were most stable across AUC and F1-score for all backbones; YOLOv7 AUC rose from 77.3% to 79.2% (+2.4%) and F1-score from 84.9% to 87.3% (+2.7).ResNet50 achieved 77.7% AUC and 90.2% F1-score; InceptionV3 achieved 86.1% AUC and 89.3% F1-score.Source: Figure 3 · p. 4, §4.1 · p. 4
- On DS3, InceptionV3 AUC rose from 71.4% to 93.3% (+30.7), the clearest reported gain; harmonization also produced the strongest, most consistent results across backbones.The comparison includes AUC and F1-score across YOLOv7, ResNet50, and InceptionV3.Source: Figure 4 · p. 4, §4.1 · p. 4
- On DS3, YOLOv7 sensitivity increased from 42.7% to 63.1% (47.7%); InceptionV3 specificity rose from 20.0% to 50.0% (+150%) while sensitivity reached 97.0%.Source: §4.1 · p. 4
4.2. Capsule endoscopy
- On DS5, harmonization delivered the strongest AUC across all backbones and the best F1-score for ResNet50 and InceptionV3; FDA led YOLOv7 F1-score, where harmonization reached 88.8%.YOLOv7 AUC rose from 75.6% to 85.13% (+12.6%).Source: Figure 5 · p. 5, §4.2 · p. 5
5. CONCLUSION
The proposed lightweight harmonization consistently improved cross-dataset generalization for detection and classification under a strict train-on-one, test-on-others protocol. It uses fixed reference-domain CIE-Lab statistics without parameter tuning; future work includes federated learning and other color-imaging modalities.
5. CONCLUSION
- Harmonization consistently improved cross-dataset generalization for detection and classification under a strict train-on-one, test-on-others protocol.The study kept architectures, trained weights, inference, and evaluation settings unchanged.Source: §5 · p. 5
- The method achieved the best results in most dataset–backbone combinations and a more stable profile across AUC, F1-score, sensitivity, and specificity.Advanced baselines were competitive in some cases but required method-specific parameter choices.Source: §5 · p. 5
- Fixed reference-domain CIE-Lab statistics avoid parameter tuning, making the preprocessing simpler and more reproducible for cross-dataset endoscopic analysis.The conclusion identifies small target datasets and situations where retraining is undesirable as practical use cases.Source: §5 · p. 5
- Future work includes client-side normalization in federated learning and extensions to histopathology and retinal imaging.Source: §5 · p. 5