Influence of color correction on pathology detection in Capsule Endoscopy

TL;DR

This study investigates how color correction impacts pathology detection in Wireless Capsule Endoscopy by benchmarking RetinaNet and YOLOv5 on three data variants: the original SEE-AI dataset and two color-corrected versions produced with CC and CCC. Across these variants, color correction tends to enlarge predicted bounding boxes and increase false positives, yet it does not yield consistent improvements in key detection metrics such as F1, IoU, or AP50, with effects varying by pathology. The analysis highlights a trade-off between color fidelity and diagnostic accuracy, noting that color correction can reduce image contrast and may be more beneficial when paired with contrast enhancement. The work provides datasets, code, and a nuanced view of preprocessing choices for WCE pathology detection, guiding future efforts to integrate color correction with contrast strategies.

Abstract

Pathology detection in Wireless Capsule Endoscopy (WCE) using deep learning has been explored in the recent past. However, deep learning models can be influenced by the color quality of the dataset used to train them, impacting detection, segmentation and classification tasks. In this work, we evaluate the impact of color correction on pathology detection using two prominent object detection models: Retinanet and YOLOv5. We first generate two color corrected versions of a popular WCE dataset (i.e., SEE-AI dataset) using two different color correction functions. We then evaluate the performance of the Retinanet and YOLOv5 on the original and color corrected versions of the dataset. The results reveal that color correction makes the models generate larger bounding boxes and larger intersection areas with the ground truth annotations. Furthermore, color correction leads to an increased number of false positives for certain pathologies. However, these effects do not translate into a consistent improvement in performance metrics such as F1-scores, IoU, and AP50. The code is available at https://github.com/agossouema2011/WCE2024. Keywords: Wireless Capsule Endoscopy, Color correction, Retinanet, YOLOv5, Detection

Figures (7)

• Figure 1: SEE-AI Dataset Annotations per class. We can observe that classes like lymph-follicle, erosion and polyp-like have more annotations, where as SMT, stenosis, and diverticulum have very few annotations.
• Figure 3: Retinanet F1-scores across different color schemes. The F1 scores increase with both color corrections (CC and CCC) for angiodysplasia, lymphangiectasia and vein. Only CC increases F1-score for SMT while CCC increases F1-score polyp-like.
• Figure 4: YOLOv5 F1-scores across different color schemes. The F1 scores increase with both color corrections only for lymphangiectasia. Only CC increases F1-score for stenosis while CCC increases F1-score for angiodysplasia.
• Figure 5: Retinanet AP across different color schemes. The AP increase with both color corrections (CC and CCC) for angiodysplasia, SMT and polyp-like. Only CC increases AP for Erythema while CCC increases AP for bleeding and vein.
• Figure 6: YOLOv5: AP across different color schemes. The AP increase with both color corrections (CC and CCC) for lymphangiectasia, polyp-like and bleeding. Only CC increases AP for stenosis, while CCC increases AP for erosion and angiodysplasia.