TRUE-Colon: Exposing a Consistent Transfer Asymmetry in Real-Time Polyp Detection

Sebastian Doerrich; Andreas Franz Schwab; Francesco Di Salvo; Shyam Nandan Rai; Hanh Huyen My Nguyen; Christian Ledig

eess.IVAug 13, 2026

TRUE-Colon: Exposing a Consistent Transfer Asymmetry in Real-Time Polyp Detection

Authors: Sebastian Doerrich, Andreas Franz Schwab, Francesco Di Salvo, Shyam Nandan Rai, Hanh Huyen My Nguyen, Christian Ledig

Organizations: XAI Lab Bamberg, University of Bamberg, Bamberg, Germany

Abstract

Computer-aided detection (CADe) systems for colonoscopy promise to reduce clinical miss rates, yet reliable real-world deployment remains elusive. This translational gap stems in part from a structural flaw in model development: the reliance on curated datasets that under-represent the long negative stretches and procedure-related artifacts characteristic of routine examinations. Training and evaluating architectures strictly on these lesion-centric benchmarks creates an illusion of success, since such benchmarks cannot capture clinically crucial metrics. To expose this gap, we establish TRUE-Colon, a standardized benchmarking protocol that measures key deployment characteristics alongside localization accuracy, and evaluate four real-time architectures (Faster R-CNN, YOLOv8, YOLOv11, RT-DETR) across curated benchmarks (SUN, PICCOLO) and 60 unedited, full-length procedures (REAL-Colon). We observe a consistent transfer asymmetry: models trained strictly on curated clips suffer a severe performance collapse when evaluated on full procedures, whereas procedure-trained models substantially improve rejection of non-polyp content on REAL-Colon, and largely retain their accuracy on curated benchmarks. Beyond transferability, we find that the Transformer detector attains the strongest sensitivity and the earliest, most persistent detections, while the convolutional detectors stay competitive at a higher throughput. Together, these results indicate that both training and benchmarking for deployable CADe should shift from curated, lesion-centric clips toward full-procedure data and deployment-relevant operating points. Source code is available at https://github.com/sdoerrich97/true-colon.

Figures & tables

Panel of Figure 1 — Figure 1: Qualitative YOLOv11 results on representative colonoscopy frames. The examples illustrate varying levels of difficulty, including clear polyp visibility, specular highlights, motion blur, and challenging background artifacts. Green bounding boxes show the ground-truth, while red bounding boxes indicate predictions.

Dataset	Split	POS	NEG	Total	Neg (%)
REAL-Colon [ 1 ]	Train	230,935	1,475,700	1,706,635	86.47
(CC BY 4.0)	Val	29,067	330,236	359,303	91.91
	Test	82,106	489,979	572,085	85.65
SUN [ 15 ]	Train	34,915	58,839	93,754	62.76
(Research Agreement)	Val	5,792	23,845	29,637	80.46
	Test	8,429	26,869	35,298	76.12

Table 1: Dataset distribution including positive/negative frame counts per split and proportion of negative frames. REAL-Colon maintains the highest proportion of negative frames (

>85\%

) in contrast to the curated benchmarks SUN and PICCOLO.

Dataset (Train $\rightarrow$ Test)	Model	$\tau^{\star}$	mAP 50:95	mAP 50	TPR	FPR
SUN $\rightarrow$ SUN	Faster R-CNN	0.07	0.408	0.738	0.739	0.044
	YOLOv8	0.07	0.396	0.694	0.731	0.046
	YOLOv11	0.10	0.411	0.724	0.735	0.047
	RT-DETR	0.12	0.389	0.689	0.725	0.050
PICCOLO $\rightarrow$ PICCOLO	Faster R-CNN	–	0.531	0.743	–	–
	YOLOv8	–	0.487	0.673	–	–

Table 2: Curated in-domain baselines on SUN and PICCOLO. Frame-level TPR/FPR are reported for SUN only; PICCOLO frame-level metrics are omitted because the official test split contains only one negative image. The convolutional detectors achieve high accuracy on both curated domains, while RT-DETR underperforms on PICCOLO.

Curated $\rightarrow$ Procedure (Test: REAL-Colon)
Dataset (Train $\rightarrow$ Test)	Model	$\tau^{\star}$	mAP 50:95	mAP 50	TPR	FPR
SUN $\rightarrow$ REAL-Colon	Faster R-CNN	0.20	0.131	0.220	0.290	0.043
	YOLOv8	0.25	0.105	0.177	0.261	0.045
	YOLOv11	0.04	0.097	0.164	0.240	0.040
	RT-DETR	0.25	0.132	0.225	0.310	0.040
PICCOLO $\rightarrow$ REAL-Colon	Faster R-CNN	0.99	0.080	0.154	0.271	0.060

Table 3: Cross-distribution transfer between curated benchmarks (SUN, PICCOLO) and full procedures (REAL-Colon). For curated

\rightarrow

REAL-Colon, thresholds are chosen to approximately match the evaluated frame-level FPR; reverse transfer reports detection-level generalization back to curated domains.

	Metric	F. R-CNN	YOLOv8	YOLOv11	RT-DETR
	$\text{mAP}_{50}$ ( $\uparrow$ )	$0.324{\scriptstyle\pm 0.03}$	$0.413{\scriptstyle\pm 0.00}$	$0.384{\scriptstyle\pm 0.01}$	$\mathbf{0.488{\scriptstyle\pm 0.02}}$
A: Det.	$\text{mAP}_{50{:}95}$ ( $\uparrow$ )	$0.196{\scriptstyle\pm 0.02}$	$0.272{\scriptstyle\pm 0.01}$	$0.245{\scriptstyle\pm 0.01}$	$\mathbf{0.327{\scriptstyle\pm 0.01}}$
	TPR $\tau$ ( $\uparrow$ )	$0.463{\scriptstyle\pm 0.06}$	$0.508{\scriptstyle\pm 0.02}$	$0.453{\scriptstyle\pm 0.00}$	$\mathbf{0.720{\scriptstyle\pm 0.04}}$
	FPR $\tau$ ( $\downarrow$ )	$0.046{\scriptstyle\pm 0.02}$	$0.020{\scriptstyle\pm 0.00}$	$\mathbf{0.017{\scriptstyle\pm 0.00}}$	$0.082{\scriptstyle\pm 0.03}$
	TPR $\tau^{\star}$ ( $\uparrow$ )	$0.463{\scriptstyle\pm 0.07}$	$0.605{\scriptstyle\pm 0.03}$	$0.573{\scriptstyle\pm 0.01}$	$\mathbf{0.651{\scriptstyle\pm 0.05}}$
B: Frame	FPR $\tau^{\star}$ ( $\downarrow$ )	$0.046{\scriptstyle\pm 0.02}$	$0.044{\scriptstyle\pm 0.01}$	$0.043{\scriptstyle\pm 0.00}$	$\mathbf{0.043{\scriptstyle\pm 0.02}}$

Table 4: REAL-Colon test performance at a fixed (

\tau=0.2

) and a normalized (

\tau^{\star}

targeting

\mathrm{FPR}\approx 4

–

5\%

, selected post-hoc on the test split for cross-model normalization) threshold. Bold marks the best mean per row; lesion-level differences (block C,

n=21

) fall within seed variance and are not claimed as significant.

Appendix figures & tables17 assets

Supplementary material from the paper’s appendix.

Appendix

Video	Range	Video	Range	Video	Range
1_1	920–45137	1_2	1620–25974	1_3	1967–39372
1_4	850–45385	1_5	360–28544	1_6	786–48217
1_7	1050–37920	1_8	1165–34891	1_9	2031–39884
1_10	2075–46215	1_11	380–23390	1_12	530–40842
1_13	1996–22417	1_14	347–28711	1_15	910–30831
2_1	1–25421	2_2	414–24916	2_3	1–47191

Appendix

Table 0.B1: Retained inside-patient frame-ID intervals for the 60 REAL-Colon videos (labeled patient_video). Frames outside the listed range are discarded as out-of-patient.

Cases 1–57				Cases 58–113
Case ID	Type	Split	#Images	Case ID	Type	Split	#Images
1	POS	train	527	58	POS	train	267
2	POS	train	1313	59	POS	train	646
3	POS	train	292	60	POS	train	146
4	POS	train	80	61	POS	train	679
5	POS	train	930	62	POS	train	351

Appendix

Table 0.B2: SUN case-level image counts and train/val/test assignment (positive cases IDs 1–100, negative cases 101–113). Splitting by case rather than by frame prevents patient-level leakage into the evaluation folds.

	Faster R-CNN	YOLOv8	YOLOv11	RT-DETR
$\mathrm{mAP}_{50}$	$0.324{\scriptstyle\pm 0.032}$	$0.413{\scriptstyle\pm 0.002}$	$0.384{\scriptstyle\pm 0.009}$	$0.488{\scriptstyle\pm 0.019}$
$\mathrm{mAP}_{75}$	$0.216{\scriptstyle\pm 0.023}$	$0.299{\scriptstyle\pm 0.010}$	$0.263{\scriptstyle\pm 0.005}$	$0.359{\scriptstyle\pm 0.009}$
$\mathrm{mAP}_{50:95}$	$0.196{\scriptstyle\pm 0.021}$	$0.272{\scriptstyle\pm 0.005}$	$0.245{\scriptstyle\pm 0.005}$	$0.327{\scriptstyle\pm 0.009}$
$\mathrm{AR}_{1}$	$0.295{\scriptstyle\pm 0.014}$	$0.376{\scriptstyle\pm 0.003}$	$0.357{\scriptstyle\pm 0.005}$	$0.415{\scriptstyle\pm 0.011}$
$\mathrm{AR}_{10}$	$0.387{\scriptstyle\pm 0.030}$	$0.466{\scriptstyle\pm 0.009}$	$0.453{\scriptstyle\pm 0.012}$	$0.585{\scriptstyle\pm 0.007}$
$\mathrm{AR}_{100}$	$0.394{\scriptstyle\pm 0.035}$	$0.471{\scriptstyle\pm 0.012}$	$0.457{\scriptstyle\pm 0.013}$	$0.672{\scriptstyle\pm 0.008}$

Appendix

Table 0.C1: Extended detection-level performance on REAL-Colon (mean

\pm

std over three seeds).

\mathrm{mAP}_{t}

denotes mean average precision at IoU threshold

t

\mathrm{mAP}_{50:95}

averages over IoU

0.50

–

0.95

, and

\mathrm{AR}_{k}

is average recall allowing

k

detections per image. RT-DETR leads at every threshold and recall budget.

	Faster R-CNN	YOLOv8	YOLOv11	RT-DETR
$\mathrm{AP}_{\text{small}}$	$0.000$	$0.000$	$0.000$	$0.002$
$\mathrm{AP}_{\text{medium}}$	$0.012{\scriptstyle\pm 0.006}$	$0.049{\scriptstyle\pm 0.009}$	$0.054{\scriptstyle\pm 0.005}$	$0.044{\scriptstyle\pm 0.014}$
$\mathrm{AP}_{\text{large}}$	$0.207{\scriptstyle\pm 0.021}$	$0.284{\scriptstyle\pm 0.005}$	$0.255{\scriptstyle\pm 0.005}$	$0.343{\scriptstyle\pm 0.010}$

Appendix

Table 0.C2: Average precision by object size on REAL-Colon, following the COCO definition (small

{\leq}32^{2}

, medium

32^{2}

–

96^{2}

, large

{\geq}96^{2}

pixels), as mean

\pm

std over three seeds. Small-polyp AP is near zero for all detectors, so localization is driven by medium and large lesions.

Panel of Figure 0 — Figure 0.C1: AFROC curves on REAL-Colon. Left: model-specific confidence thresholds ( $\mathrm{FPR}\approx 4$ – $5\%$ ). Right: fixed $\tau=0.2$ for all models. The $y$ -axis is the fraction of ground-truth lesion boxes detected and the $x$ -axis is the fraction of negative frames raising at least one false-positive alert. RT-DETR retains the highest sensitivity at the lowest false-positive fractions.

Shared confidence $\tau=0.20$
Metric	Faster R-CNN	YOLOv8	YOLOv11	RT-DETR
Precision	$0.524{\scriptstyle\pm 0.086}$	$0.701{\scriptstyle\pm 0.022}$	$0.715{\scriptstyle\pm 0.011}$	$0.398{\scriptstyle\pm 0.095}$
Recall	$0.463{\scriptstyle\pm 0.066}$	$0.508{\scriptstyle\pm 0.022}$	$0.453{\scriptstyle\pm 0.004}$	$0.720{\scriptstyle\pm 0.041}$
Model-specific thresholds (FPR $\approx$ 4–5%)
Precision	$0.524{\scriptstyle\pm 0.086}$	$0.523{\scriptstyle\pm 0.029}$	$0.516{\scriptstyle\pm 0.017}$	$0.577{\scriptstyle\pm 0.094}$
Recall	$0.463{\scriptstyle\pm 0.066}$	$0.605{\scriptstyle\pm 0.030}$	$0.573{\scriptstyle\pm 0.009}$	$0.651{\scriptstyle\pm 0.049}$

Appendix

Table 0.C3: Detection-level precision and recall under three confidence-threshold strategies (shared

\tau=0.2

, model-specific, and per-seed; mean

\pm

std over three seeds). A shared threshold penalizes RT-DETR, which over-fires at low

\tau

; model-specific thresholds place every detector at a comparable false-positive rate.

Model-specific thresholds (FPR $\approx$ 4–5%)
Metric	Faster R-CNN	YOLOv8	YOLOv11	RT-DETR
Sensitivity / TPR	$0.463{\scriptstyle\pm 0.066}$	$0.605{\scriptstyle\pm 0.030}$	$0.573{\scriptstyle\pm 0.009}$	$0.651{\scriptstyle\pm 0.049}$
Specificity	$0.954{\scriptstyle\pm 0.017}$	$0.956{\scriptstyle\pm 0.006}$	$0.957{\scriptstyle\pm 0.003}$	$0.957{\scriptstyle\pm 0.017}$
FPR	$0.046{\scriptstyle\pm 0.017}$	$0.044{\scriptstyle\pm 0.006}$	$0.043{\scriptstyle\pm 0.003}$	$0.043{\scriptstyle\pm 0.017}$
Precision	$0.637{\scriptstyle\pm 0.086}$	$0.698{\scriptstyle\pm 0.023}$	$0.693{\scriptstyle\pm 0.020}$	$0.723{\scriptstyle\pm 0.070}$
$F_{1}$ score	$0.532{\scriptstyle\pm 0.047}$	$0.648{\scriptstyle\pm 0.015}$	$0.627{\scriptstyle\pm 0.013}$	$0.681{\scriptstyle\pm 0.001}$

Appendix

Table 0.C4: Frame-level performance under the model-specific and per-seed threshold strategies (mean

\pm

std over three seeds), extending the frame-level block of Table 4 with specificity, precision,

F_{1}

, and

F_{2}

. At matched operating points RT-DETR attains the best

F_{1}

and

F_{2}

. Per-seed refitting only reduces variance.

Fixed confidence threshold $\tau=0.20$
	Faster R-CNN	YOLOv8	YOLOv11	RT-DETR
Lesions detected ( $\geq 1$ match)	21.0	21.0	21.0	21.0
Lesions detected ( $\geq 25\%$ frames)	19.0	19.3	18.7	21.0
Lesions detected ( $\geq 50\%$ frames)	11.7	12.7	12.7	19.7
Detected within 1 s	1 1.0	1 1.7	1 1.0	1 7.7
Detected within 3 s	1 7.0	1 5.3	1 5.3	12.0

Appendix

Table 0.C5: Lesion-level consistency and latency on REAL-Colon (

n=21

lesions, means over three seeds), under a fixed

\tau=0.20

and under model-specific thresholds. Rows report how many lesions are detected at all and on

{\geq}25\%

{\geq}50\%

of their visible frames, how many are first detected within

1

3

5

s, and the latency to the first detected frame. RT-DETR detects lesions more persistently and earlier than the convolutional detectors.

Detectron2
Model	Inf [ms]	Eval [ms]	Total [ms]	FPS ${}_{\text{inf}}$	FPS ${}_{\text{total}}$
Faster R-CNN	$25.07{\scriptstyle\pm 2.72}$	$0.20{\scriptstyle\pm 0.00}$	$26.70{\scriptstyle\pm 2.52}$	39.9	37.5

Appendix

Table 0.C6: Runtime on REAL-Colon at batch size 1 on one NVIDIA H100 NVL GPU (mean

\pm

std). Times are per image for pre-processing, inference, and post-processing.

\mathrm{FPS}_{\text{inf}}

counts inference only and

\mathrm{FPS}_{\text{total}}

the full pipeline. All detectors exceed the

30

FPS real-time threshold.

Resolution	conf	$\mathrm{mAP}_{50:95}$	$\mathrm{mAP}_{50}$	TPR	FPR	$\mathrm{FPS}_{\text{inf}}$	$\mathrm{FPS}_{\text{total}}$
$224\times 224$	0.01	0.184	0.289	0.444	0.044	212.9	183.4
$640\times 640$	0.05	0.246	0.384	0.584	0.043	192.3	163.9
$1024\times 1024$	0.03	0.255	0.400	0.576	0.044	184.3	143.8

Appendix

Table 0.C7: Effect of input resolution on YOLOv11-M (REAL-Colon).

640\times 640

gives the best accuracy-to-throughput trade-off and is the resolution used in the main paper.

Model	conf	$\mathrm{mAP}_{50:95}$	$\mathrm{mAP}_{50}$	TPR	FPR
Faster R-CNN	0.200	0.203	0.330	0.497	0.043
YOLOv8-M	0.013	0.221	0.342	0.524	0.048
YOLOv11-M	0.015	0.194	0.297	0.504	0.046
RT-DETR	0.250	0.290	0.432	0.634	0.050
SSD-300 (author setup)	–	0.216	0.338	0.505	0.054

Appendix

Table 0.C8: Comparison at

300\times 300

following the original REAL-Colon author setup [ 1 ] on the full dataset. Because the original IoU and confidence thresholds are unspecified, TPR and FPR are for observational reference only.

Model / Config	conf	$\mathrm{mAP}_{50:95}$	$\mathrm{mAP}_{50}$	TPR	FPR
YOLOv11-S	0.040	0.221	0.353	0.570	0.045
YOLOv11-M	0.050	0.246	0.384	0.584	0.043
YOLOv11-L	0.007	0.276	0.420	0.628	0.044
Neg. ratio 1:0.15	0.27	0.209	0.328	0.482	0.048
Neg. ratio 1:0.50	0.27	0.155	0.254	0.450	0.047
Neg. ratio 1:1.00	0.10	0.253	0.401	0.602	0.048

Appendix

Table 0.C9: Effect of model capacity, negative-sampling ratio, and optimizer on YOLOv11 (REAL-Colon,

640\times 640

unless noted). A balanced

1{:}1

negative ratio and larger capacity help most, while the optimizer barely matters.

Metric	GI Genius v1	GI Genius v2	EndoAID A	EndoAID B	EndoMind
Sensitivity / TPR	0.5063	0.6785	0.6560	0.5295	0.6022
Specificity	0.9694	0.9577	0.9719	0.9930	0.9589
FPR	0.0275	0.0380	0.0252	0.0063	0.0369
Precision	0.5966	0.5701	0.6658	0.8720	0.5499
$F_{1}$ score	0.5915	0.6391	0.6943	0.7177	0.5960
First detection (ms)	1510	607	659	1316	1083

Appendix

Table 0.C10: Frame-level performance of commercial CADe systems and the open-source EndoMind detector (

\mathrm{IoU}>0

\tau=0.2

), as reported in [ 22 ] . Values come from a different dataset and pipeline, so they contextualize rather than directly compare with our detectors.

Figure 0.C4: Challenging REAL-Colon test frames on which all evaluated architectures produce false negatives or false positives. From left to right: specular highlights, motion blur, instrument occlusion, and a flat non-polypoid lesion. These conditions are the main shared failure mode across detectors.

Figure 0.C5: Qualitative detection results across all four architectures on representative REAL-Colon test frames. Each row is one frame under varying visual conditions. The columns are Faster R-CNN, YOLOv8, YOLOv11, and RT-DETR (left to right). Green boxes are ground-truth annotations and red boxes are model predictions. Predictions agree on clear polyps and diverge on flat and NBI frames.

Medical Image Benchmarks

Medical image benchmarks compare AI methods on curated clinical images, tasks, and evaluation protocols to measure accuracy, robustness, and clinical usefulness.

See all 337 papers

DepthPolyp: Pseudo-Depth Guided Lightweight Segmentation for Real-Time Colonoscopy
May 15, 2026Zhuoyu Wu, Wenhui Ou, Lexi Zhang +5Endoscopy Image Segmentation
CyPhi AI Lab, Monash University, Malaysia Campus, Malaysia · Department of Electronic & Computer Engineering, Hong Kong University of Science & Technology, Hong Kong, P.R. China · Harbin Institute of Technology, Harbin, P.R. China +1
Metrics or Mirage? An Audit of Evaluation Inconsistencies in Colonoscopy Polyp Segmentation Benchmarks
Jul 9, 2026Aisha Urooj, Zain Ul Abdien, Neelu MadanPolyp Segmentation Benchmark Reproducibility
Lunit, South Korea · RPTU Kaiserslautern, Germany · Aalborg University & Pioneer Center, Denmark
Structure-to-Image: Zero-Shot Depth Estimation in Colonoscopy via High-Fidelity Sim-to-Real Adaptation
Feb 25, 2026Juan Yang, Yuyan Zhang, Han Jia +2Monocular Depth Estimation Synthetic-to-Real Domain Adaptation
College of Computer Science, Sichuan University, Chengdu, China · Department of Gastroenterology and Hepatology, West China Hospital, Sichuan University, Chengdu, China

May 15, 2026J/K move · Enter open · S save

DepthPolyp: Pseudo-Depth Guided Lightweight Segmentation for Real-Time Colonoscopy

Zhuoyu Wu, Wenhui Ou, Lexi Zhang, Pei-Sze Tan+4

CyPhi AI Lab, Monash University, Malaysia Campus, Malaysia · Department of Electronic & Computer Engineering, Hong Kong University of Science & Technology, Hong Kong, P.R. China · Harbin Institute of Technology, Harbin, P.R. China+1

Endoscopy Image Segmentation Domain Generalization Polyp Segmentation Medical Image Segmentation Semantic Segmentation

Open paper AskPDF arXiv

Accurate polyp segmentation in colonoscopy is essential for early colorectal cancer detection, yet real-world clinical environments pose persistent challenges such as motion blur, specular reflections, and illumination instability. Most existing methods are optimized on clean benchmark images and suffer noticeable performance degradation when deployed in authentic surgical scenarios. We propose DepthPolyp, a lightweight and robust segmentation framework based on pseudo-depth-guided multi-task learning and efficient feature modulation. The architecture combines hierarchical Ghost factorization for compact feature generation, Interleaved Shuffle Fusion for low-cost cross-scale interaction, and Dynamic Group Gating for adaptive group-wise feature weighting. Extensive experiments demonstrate that DepthPolyp achieves strong cross-dataset generalization when trained on degraded data and evaluated on both clean and noisy target domains, consistently outperforming lightweight baselines and remaining competitive with substantially larger models. In real surgical video evaluation on PolypGen, DepthPolyp achieves better segmentation performance than models up to

20\times

larger while preserving real-time inference speed. With only 3.57M parameters and 0.86 GMACs, the proposed method runs at over 180 FPS on mobile devices, making it well suited for real-time deployment in resource-constrained clinical environments. Code and pretrained weights are available at: https://github.com/ReaganWu/DepthPolyp/

Abstract

Medical Image Benchmarks

Explore similar work

DepthPolyp: Pseudo-Depth Guided Lightweight Segmentation for Real-Time Colonoscopy

Metrics or Mirage? An Audit of Evaluation Inconsistencies in Colonoscopy Polyp Segmentation Benchmarks

Structure-to-Image: Zero-Shot Depth Estimation in Colonoscopy via High-Fidelity Sim-to-Real Adaptation