Source-linked AI summary
Colonoscopy Polyp Detection and Classification: Dataset Creation and Comparative Evaluations
Kaidong Li, Mohammad I. Fathan, Krushi Patel, Tianxiao Zhang, Cuncong Zhong, Ajay Bansal, Amit Rastogi, Jean S. Wang, Guanghui Wang
TL;DR
Differences among existing endoscopic datasets in image resolution and color temperature motivate a dataset that supports polyp detection and classification. The authors assemble and annotate such a dataset and compare eight detectors, reporting that several models perform relatively well across tests.
Problem
Existing endoscopic datasets differ in resolution and color temperature, while image-background and quality variation are identified as challenges for real-world CNN performance.
Source: Dataset Build · p. 11Method
The authors combine publicly available and newly collected endoscopic data, standardize annotations, label polyp classes and frame-level bounding boxes with endoscopists, and evaluate eight detectors.
Source: Dataset Build · p. 12, Datasets Selection and Annotation · p. 12Results
YOLOv4, ATSS, and RefineDet perform relatively well across tests with balanced precision and recall, while the one-class detector reaches 91% adenomatous-polyp detection sensitivity.
Source: Conclusion · p. 22Takeaways & Limitations
The dataset is proposed as a standardized platform for training and evaluating polyp-classification models and as a baseline for future research.
Source: Conclusion · p. 22
Abstract
from arXiv · showhide
Colorectal cancer (CRC) is one of the most common types of cancer with a high mortality rate. Colonoscopy is the preferred procedure for CRC screening and has proven to be effective in reducing CRC mortality. Thus, a reliable computer-aided polyp detection and classification system can significantly increase the effectiveness of colonoscopy. In this paper, we create an endoscopic dataset collected from various sources and annotate the ground truth of polyp location and classification results with the help of experienced gastroenterologists. The dataset can serve as a benchmark platform to train and evaluate the machine learning models for polyp classification. We have also compared the performance of eight state-of-the-art deep learning-based object detection models. The results demonstrate that deep CNN models are promising in CRC screening. This work can serve as a baseline for future research in polyp detection and classification.
Motivation
CRC can develop from precancerous polyps over several years, making early screening important. The paper motivates colonoscopy assistance and a labeled dataset to support polyp detection and classification.
Motivation
- CRC develops from precancerous polyps over years, creating an opportunity for regular screening before cancer develops.The passage says development can take several years, even up to a decade.Source: Motivation · p. 2
CRC screening options
- Screening tests balance sensitivity against specificity, and the paper describes sensitivity as especially important because missed lesions can have severe consequences.A higher-specificity test may follow a high-sensitivity test to filter false positives.Source: CRC screening options · p. 2
- Colonoscopy is the recommended visual screening method, with high sensitivity and lesion removal at detection, but preparation and procedure burdens can reduce compliance.The passage also notes sedation or anesthesia and a risk of post-colonoscopy bleeding.Source: CRC screening options · p. 2
Goals
- Colonoscopy performance can vary with preparation, polyp characteristics, physician experience, and fatigue; one cited study found 20% more polyps in early-morning screenings.The passage reports lower performance before training than after training.Source: Goals · p. 3
- The proposed computer-aided framework is intended to detect and classify polyps in real time and prompt physicians to their regions and categories.The passage describes assistance with missed detections associated with fatigue and clinics lacking experienced gastroenterologists.Source: Goals · p. 3
- The authors build a video-based polyp dataset with frame-level locations and categories, then use it to evaluate state-of-the-art detection models.The dataset is intended to include different polyp stages and types.Source: Goals · p. 4
Related Work
The paper situates polyp detection within medical computer vision and focuses on deep-learning object detectors that jointly produce object categories and locations.
Computer Vision in Medical Applications
- Deep CNNs extract high-level features, while earlier medical computer-vision work used handcrafted features and models for tasks including polyp detection and classification.The paper distinguishes its focus on detection in endoscopic sequences from prior classification work that assumed polyps were already detected and cropped.Source: Computer Vision in Medical Applications · pp. 4–5
Object Detection
- The study focuses on object detection, which returns object classes and bounding-box locations, rather than pixel-level segmentation.The passage says segmentation is usually more time-consuming and pixel-level classification is not necessary for this task.Source: Object Detection · p. 5
- Two-stage detectors separate region proposal and classification, generally offering higher accuracy but more processing time; one-stage detectors combine the stages and often achieve real-time performance.The passage says one-stage accuracy has been catching up and can now be comparable.Source: Object Detection · p. 5
Evaluation Models for Detection and Classification
The paper compares eight object-detection models, spanning one- and two-stage detectors and varied strategies for feature extraction, anchor refinement, and sample selection.
Evaluation Models for Detection and Classification
- The evaluation covers eight state-of-the-art object detection and classification models, including Faster R-CNN, YOLO variants, SSD, RetinaNet, DetNet, RefineDet, and ATSS.The introduction describes the study as a comparative evaluation of these models.Source: Evaluation Models for Detection and… · pp. 5–7
- Faster R-CNN shares backbone features between its region proposal network and classification module, while its trainable proposal network replaces selective search.The RPN regresses box locations and objectness scores from anchors at sliding-window positions.Source: Evaluation Models for Detection and… · pp. 5–6, Figure 1 · p. 6
- YOLOv4 combines selected detection-model techniques with CSPDarknet53, spatial pyramid pooling, and PANet while aiming to retain high-speed inference.The passage describes the selected techniques as a way to enhance performance without sacrificing too much inference cost.Source: Evaluation Models for Detection and… · p. 7
- SSD predicts class scores and box offsets from multiple feature-map scales, using anchor boxes to support detection across object sizes.Its simple one-stage framework is described as fast and as balancing speed with accuracy.Source: Evaluation Models for Detection and… · pp. 7–8, Figure 2 · p. 7
- RetinaNet builds on SSD with a feature pyramid and focal loss, which assigns greater weight to difficult foreground objects than easy background cases.The passage identifies class imbalance as especially pronounced in one-stage models.Source: Evaluation Models for Detection and… · p. 8
- DetNet preserves feature-map resolution in its later stages and uses dilated convolutions to increase receptive field, and this study applies it as a Faster R-CNN backbone.The design addresses the loss of small objects and localization accuracy associated with progressively lower resolutions.Source: Evaluation Models for Detection and… · p. 8, Figure 3 · p. 9
- RefineDet uses ARM to filter and coarsely adjust anchors, TCB to combine deeper semantic features with shallower features, and ODM to refine boxes and predict classes.The two-step box refinement and multi-level information are described as improving results.Source: Evaluation Models for Detection and… · pp. 8–9, Figure 4 · p. 10
- ATSS selects positive and negative training samples dynamically using candidate-anchor IoU statistics and whether candidate centers lie inside the ground-truth box.The selection threshold is the candidate IoU mean plus its standard deviation.Source: Evaluation Models for Detection and… · pp. 9, 11
Dataset Build
The paper combines endoscopy datasets to address variation in imaging conditions and a shortage of distinct polyp examples, then annotates and organizes the data as a benchmark for detection and classification. Sampling, frame curation, and sequence-level splitting shape the resulting dataset.
Dataset Build
- Existing collections vary in image resolution and color, while repeated viewpoints leave many sequences with redundant frames and few distinct polyps.These differences limit the representativeness of the available data for training and evaluating classifiers across imaging sources.Source: Dataset Build · pp. 11–12
- The combined dataset standardizes annotations for hyperplastic and adenomatous polyps, whose classifications have different clinical implications.It combines publicly available datasets with a new University of Kansas Medical Center dataset.Source: Dataset Build · p. 12, Datasets Selection and Annotation · p. 12
Datasets Selection and Annotation
- Three endoscopists labeled polyp classes for collected sequences and bounding boxes in every frame; existing datasets lacking labels were supplemented or converted.MICCAI 2017 and CVC colon DB supplied masks but no classification labels, while GLRC lacked polyp-location labels.Source: Datasets Selection and Annotation · pp. 12–13
Frame Selection
- Adaptive sampling uses camera movement and sequence length to reduce redundancy, retaining all frames from short sequences and about 300–500 frames from long ones.Heavily blurred, out-of-focus, or illumination-shifted frames were removed, while less flawed frames were retained for robustness; the final dataset contains 155 sequences and 37,899 frames.Source: Frame Selection · p. 13
Dataset Split
- Splitting by video sequence avoids placing frames of the same polyp in training, validation, and test sets; the resulting sets contain 116, 17, and 22 sequences, respectively.Each dataset and polyp class was split independently at 75%, 10%, and 15%; the dataset is publicly accessible.Source: Dataset Split · p. 14
Experiments
The experiments compare eight object detectors across frame-based detection and sequence-based classification setups. They also describe how each task is evaluated and why frame- and sequence-level predictions address different screening needs.
Experiments
- Eight object detectors are evaluated in three setups: frame-based two-class detection, frame-based one-class detection, and sequence-based two-class classification.Precision, recall, and F-score are used to evaluate model performance.Source: Experiments · pp. 14–15
Frame-based Two-class Polyp Detection
- Frame-based two-class detection evaluates each frame individually for both localization and classification.The models are trained on frames mixed across video sequences, and the setup is motivated by potential real-time assistance during colonoscopy.Source: Frame-based Two-class Polyp Detection · p. 16
Frame-based One-class Polyp Detection
- Frame-based one-class detection treats hyperplastic and adenomatous polyps as a single class to test localization without category classification.The setup reflects the stated priority of detecting polyps for follow-up screening and diagnosis.Source: Frame-based One-class Polyp Detection · p. 16
Sequence-based Two-class Polyp Classification
- Sequence-based two-class classification assigns each video the most frequently predicted category across its frames.Frames in a sequence show the same polyp from different viewpoints, which can reduce the influence of occluded or distant views.Source: Sequence-based Two-class Polyp… · pp. 16–17
Results and Analysis
Across frame-based detection, RefineDet led combined two-class performance, while detector performance varied by task and training data. Sequence-based classification results also differed in both category accuracy and consistency across video frames.
Results and Analysis
- Models were selected by validation performance, and precision, recall, and F1 were calculated at confidence 0.5; most best CNN models appeared before epoch 10.RefineDet one-class detection was an exception: 88.05% mAP at 30k iterations was similar to 88.12% at 130k iterations.Source: Results and Analysis · p. 17
Frame-based Two-class Polyp Detection
- RefineDet achieved the highest mean F1-score, mAP, and mean recall in two-class detection; YOLOv3 had the best precision but abnormally low recall.The comparison covers all evaluated detectors; adenomatous polyps were easier to detect than hyperplastic polyps. SSD reached 67.6% mAP versus 57.7% for Faster RCNN.Source: Frame-based Two-class Polyp Detection · pp. 17, 19, Table 3 · p. 18
Generalizability and Comparison with Previous Dataset
- Training only on KUMC reduced average performance by 8% on the combined test set, with representativeness, sample count, and dataset color and illumination differences cited as factors.KUMC-only training was evaluated against the full combined test set.Source: Generalizability and Comparison with… · p. 20, Table 4 · p. 20
Frame-based One-class Polyp Detection
- Detection-only results favored YOLOv3 in precision, while RefineDet remained strongest overall in F1 and AP; DetNet improved precision by 20%+ over Faster RCNN, yielding a better F1 score.The authors suggest refined location information may matter more for distinguishing polyp categories than for locating polyps.Source: Frame-based One-class Polyp Detection · pp. 20, 22, Table 5 · p. 21
- Faster RCNN achieved the best one-class recall, including 93.3% for adenomatous polyps; RefineDet led overall with the highest F1 score and AP.Faster RCNN, RefineDet, and ATSS each exceeded 80% recall for hyperplastic polyps.Source: Frame-based One-class Polyp Detection · p. 20, Table 6 · p. 21
Sequence-based Two-class Polyp Classification
- SSD, DetNet, and YOLOv4 stood out in sequence-based classification precision, recall, and F1, while RefineDet, SSD, YOLOv4, and ATSS were more consistent across sequences.The latter models had most sequences above 70% dominant-class predictions; some other detectors were close to 50% on certain sequences.Source: Table 7 · p. 22, Sequence-based Two-class Polyp… · p. 22
Frame-based One-class Polyp Detection
- In one-class detection, single-stage detectors were faster than two-stage detectors; SSD and YOLOv3 ran at 17 ms, while DetNet ran at 64 ms.The models were evaluated on an NVIDIA TESLA P100 GPU; the paper reports these times as above 60 fps and 15 fps, respectively.Source: Table 5 · p. 21, Frame-based One-class Polyp Detection · p. 22
Conclusion
The paper presents a relatively large endoscopic dataset and compares eight deep-learning object detectors for polyp detection and classification. The authors conclude that deep CNNs are promising for CRC screening and that the dataset can support future research.
Conclusion
- The authors developed a relatively large endoscopic dataset and evaluated eight state-of-the-art deep-learning object detectors.They report adenomatous-polyp sensitivity of 91% with a one-class detector and around 70% precision in classification.Source: Conclusion · p. 22
- YOLOv4, ATSS, and RefineDet performed relatively well across tests, with balanced precision and recall and consistent results for the same lesions.The authors also report that refining location information before classification effectively boosts performance.Source: Conclusion · p. 22
- The dataset is presented as a standardized platform and baseline for future polyp detection and classification research.The authors say it can help researchers design more specialized CNN models for polyp classification.Source: Conclusion · pp. 22, 24