Negative Samples for Improving Object Detection—A Case Study in AI-Assisted Colonoscopy for Polyp Detection

Abstract

Deep learning object-detection models are being successfully applied to develop computer-aided diagnosis systems for aiding polyp detection during colonoscopies. Here, we evidence the need to include negative samples for both (i) reducing false positives during the polyp-finding phase, by including images with artifacts that may confuse the detection models (e.g., medical instruments, water jets, feces, blood, excessive proximity of the camera to the colon wall, blurred images, etc.) that are usually not included in model development datasets, and (ii) correctly estimating a more realistic performance of the models. By retraining our previously developed YOLOv3-based detection model with a dataset that includes 15% of additional not-polyp images with a variety of artifacts, we were able to generally improve its F1 performance in our internal test datasets (from an average F1 of 0.869 to 0.893), which now include such type of images, as well as in four public datasets that include not-polyp images (from an average F1 of 0.695 to 0.722).

Optical diagnosis in still images of colorectal polyps: comparison between expert endoscopists and PolyDeep, a Computer-Aided Diagnosis system

“…PolyDeep is an artificial intelligence CADe/x system for the detection and characterization of colorectal polyps ( 3 , 9 , 11 , 12 ). This system is composed of two DL models, capable of detecting and classifying polypoid lesions in real time during colonoscopy ( 11 ).…”

“…Given the clinical interest in the prevention of lower gastrointestinal disease and the increasing adoption of CADe/x systems, we decided to perform an in vitro analysis on still images of colorectal polyps to compare the optical diagnosis of expert endoscopists and PolyDeep, a CADe/x system developed by our research group in previous works ( 3 , 9 , 11 , 12 ).…”

REAL-Colon: A dataset for developing real-world AI applications in colonoscopy

“…In contrast, the reality of colonoscopy videos features 80–90% negative frames (as also illustrated in Fig. 4 ), which are important for realistic AI model benchmarking and training as outlined by recent literature 23 , 26 . Furthermore, video frames from public datasets are seldom not at native spatial-temporal resolution and typically sourced from a limited number of centers.…”

“…The discrepancy between the available datasets and the real-world scenario inevitably affects both the design and development of CAD(e/x) algorithms, with a large part of academic research works to date still focusing on frame-by-frame approaches placing little emphasis on live processing speed and latency or full-procedure evaluation. Thus resulting in sub-optimal learning and unrealistic AI model performance assessments 13 , 23 , 26 , 28 . Similarly, open research challenges 18 , 28 – 32 , primarily centered on the accuracy of polyp detection, segmentation, and classification tasks, have gradually shifted their focus towards enhancing model robustness, speed and efficiency.…”