Bridging innovation and clinical reality: Interpreting the comparative study of deep learning models for multi-class upper gastrointestinal disease segmentation

Table 1 Comparative model summary

Model name	Architecture family	Pretraining	Mean IoU, %	Mean Dice	PET score, %	GRR, %	Inference latency (milliseconds/frame)	Model size	Notes
U-Net	CNN	N/A	74.88	(Self 79.37 + EDD2020 67.63)/2 = 73.50%	45.74	65.41	3.64	31.46 M, approximately 126 MB	Classic encoder-decoder with skip connections; excellent throughput with very high FPS but lower PET; simple architecture may limit generalization
ResNet + U-Net	CNN	ResNet encoder typically ImageNet-1K	80.63	(83.59 + 73.97)/2 = 78.78%	82.58	67.03	5.18	32.52 M, approximately 130 MB	Residual backbone improves accuracy and robustness with modest compute; good PET
ConvNeXt + UPerNet	CNN	Typically ImageNet-1K	82.69	(85.65 + 76.90)/2 = 81.27%	84.70	68.79	5.65	41.37 M, approximately 165 MB	Modern CNN with ViT-inspired design; strong accuracy and good throughput
M2SNet	CNN	N/A	80.87	(84.24 + 74.81)/2 = 79.53%	72.33	67.64	14.86	29.89 M, approximately 120 MB	Multi-scale subtraction units for improved feature complementarity and edge clarity; slower inference among CNNs
Dilated SegNet	CNN	ResNet50 backbone	80.55	(83.35 + 73.64)/2 = 78.49%	74.88	68.08	9.23	18.111 M, approximately 72 MB	Dilated convolutions for real-time polyp segmentation; good trade-off, moderate speed
PraNet	CNN	N/A	73.74	(74.15 + 61.12)/2 = 67.63%	48.81	65.79	12.16	32.56 M, approximately 130 MB	Parallel partial decoder + reverse attention for boundary refinement; decent accuracy but lower PET and generalization
SwinV2 + UPerNet	Transformer	SwinV2 typically ImageNet-pretrained	82.74	(85.59 + 76.97)/2 = 81.28%	78.41	68.12	12.19	41.91 M, approximately 168 MB	SwinV2 hierarchical transformer backbone with UPerNet decoder; strong accuracy with moderate compute
SegFormer	Transformer	Typically ImageNet-1K	82.86	(93.14 + 77.20)/2 = 85.17%	92.02	70.11	9.68	24.73 M, approximately 99 MB	Excellent performance-efficiency balance; low FLOPs (4.23 GFLOPs), good generalization: Recommended for real-time clinical use per paper
SETR-MLA	Transformer	ViT backbone	77.42	(82.14 + 71.48)/2 = 76.81%	52.45	69.67	5.55	90.77 M, approximately 363 MB	Segmentation transformer with multi-level aggregation; large parameter count but relatively fast inference in this setup
TransUNet	Hybrid (CNN + transformer)	ResNet50 + ViT	74.81	(77.35 + 65.06)/2 = 71.21%	26.39	67.14	13.03	105.00 M, approximately 420 MB	Combines CNN encoder and ViT: Strong representational power but heavy compute and lower PET
PVTV2 + EMCAD	Transformer	PVTV2 usually ImageNet-pretrained	82.91	(85.81 + 77.07)/2 = 81.44%	88.14	71.38	12.56	26.77 M, approximately 107 MB	Pyramid vision transformer v2 + efficient multi-scale decoding; strong generalization and good PET
FCBFormer	Transformer	PVTV2 backbone	82.00	(85.18 + 76.03)/2 = 80.61%	61.89	71.52	21.43	33.09 M, approximately 132 MB	Polyp-specialized transformer variant; strong generalization (highest GRR) but highest inference time among many models, limiting real-time use at high resolution
Swin-UMamba	Hybrid (Mamba based)	N/A	79.45	(81.23 + 71.12)/2 = 76.18%	53.31	70.93	13.00	59.89 M, approximately 240 MB	Mamba hybrid leveraging visual-state-space-model; good generalization but relatively large and slower training/inference cost
Swin-UMamba-D	Hybrid (Mamba based)	N/A	83.29	(86.15 + 77.53)/2 = 81.84%	88.39	69.36	12.97	27.50 M, approximately 110 MB	Best segmentation performance (average IoU) among study but relatively high training and inference cost; strong segmentation accuracy but moderate generalization
UMamba-Bot	Hybrid (Mamba based)	N/A	71.61	(75.03 + 61.79)/2 = 68.41%	39.45	64.78	6.27	28.77 M, approximately 115 MB	Lightweight mamba variant; good FPS but weaker accuracy and generalization
UMamba-Enc	Hybrid (Mamba based)	N/A	71.28	(75.24 + 61.82)/2 = 68.53%	37.15	65.33	7.28	27.56 M, approximately 110 MB	Encoder-focused mamba variant; similar trade-offs to UMamba-Bot: Faster but lower accuracy
VM-UNETV2	Hybrid (Mamba like)	N/A	81.63	(84.36 + 74.89)/2 = 79.63%	83.48	69.49	12.90	22.77 M, approximately 91 MB	VM-UNET encoder variant; strong PET and competitive accuracy, GPU-focused design

IoU: Intersection over union; PET: Performance-efficiency trade-off; GRR: Generalization-reliability rate; CNN: Convolutional neural network; N/A: Not applicable; FPS: Frames per second.

Table 2 Implementation roadmap

Translational challenge	Concrete solution	Responsible stakeholders	Measurable success metrics
Dataset bias and limited diversity	Multi-center data sharing agreements; standardized metadata schema (demographics, device, protocol); stratified sampling and targeted collection for underrepresented cohorts; federated learning to enable cross-site models while preserving privacy	Clinical consortiums, data governance teams, hospital IT, study PIs, legal/compliance	Number of centers and countries represented; device/vendor diversity index; demographic coverage (age/sex/ethnicity) proportions; change in GRR and external IoU on held-out sites
Annotation variability and subjectivity	Develop and enforce standardized annotation protocol and labeling guidelines; multi-expert consensus labeling; adjudication workflows; active learning to prioritize ambiguous cases; periodic re-annotation audits	Clinical experts (gastroenterologists), annotation managers, platform vendors, data scientists	Inter-rater agreement (Cohen’s kappa/mean IoU across annotators); % masks adjudicated; annotation time per case; model performance gains after consensus labels
Class imbalance/rare pathology sensitivity	Oversampling/targeted collection of rare classes; class-aware loss functions (focal, class-weighted); synthetic data and augmentation for rare classes; curriculum learning focusing on rare classes	Data acquisition teams, ML engineers, clinical partners, biostatisticians	Per-class recall/sensitivity (especially for rare classes); AUPRC for rare classes; reduction in false-negative rate for underrepresented labels
Imaging variability (lighting, specular reflection, motion blur)	Advanced preprocessing (illumination normalization, reflection removal), robust augmentation (exposure, blur, specular sim), self-supervised pretraining on large unlabeled endoscopy corpora; spatio-temporal modeling for videos	ML research team, imaging engineers, clinical endoscopy unit, vendors	Performance stratified by exposure/quality buckets (IoU under overexposed vs normal); reduction in failure cases linked to artifacts; frame-level temporal consistency metrics (temporal IoU)
Poor cross-dataset generalization/overfitting	Cross-dataset evaluation, domain adaptation techniques, federated or multi-site training, hold-out external validation sets, regularization and ensembling	ML engineers, external collaborators, validation leads, statisticians	Delta IoU/Dice between internal test and external test sets; GRR improvement on external cohorts; calibration metrics (Brier score)
Real-time performance and resource constraints (PET)	Model compression and pruning; lightweight architectures (e.g., SegFormer variants); hardware benchmark targeting (edge GPU/CPU); optimized inference pipelines	ML engineers, DevOps, clinical IT, hardware vendors	Inference latency (microseconds/frame), throughput (fps) on target hardware; memory usage, FLOPs; PET score or task-specific tradeoff metric; clinician acceptance for live use
Clinical validation and impact on workflow	Prospective clinical studies, reader studies comparing model + clinician vs clinician alone; integration pilots in endoscopy suite; user-centred UI/UX design and training	Clinical investigators, hospital operations, human factors specialists, clinical IT	Diagnostic accuracy improvement (sensitivity/specificity) in prospective trials; change in missed-lesion rate; time-to-report; clinician satisfaction and adoption rates
Trust, explainability and clinician acceptance	Provide visual explanations (attention maps, uncertainty overlays); case-level confidence scores; reporting of failure modes and limitations; clinician training modules	ML explainability team, clinical educators, product managers, regulatory/QA	Proportion of model outputs with uncertainty flags; clinician trust scores in surveys; reduction in dismissed correct alerts; explainability usability ratings
Privacy, legal & regulatory readiness	Data de-identification pipeline, DPIAs, early engagement with regulators, pre-specified validation plan, post-market surveillance plan, robust audit trails	Legal/compliance, regulatory affairs, data governance, QA, cybersecurity	Completion of DPIA and IRB approvals; regulatory submission milestones (pre-submission, submission, approvals); number of privacy incidents; time to resolve security findings
Multi-modal and longitudinal integration	Design multi-modal models (image + report + temporal video), link endoscopy frames with pathology/EMR metadata, adopt interoperable standards (DICOM/HL7/FHIR)	Data engineers, clinical informatics, pathology, ML researchers, standards officers	Increase in model performance when adding modalities (delta IoU/Dice); % cases with linked pathology; successful end-to-end FHIR/DICOM integrations; improvement in clinically-relevant outcome measures (e.g., appropriate biopsy rate)

IoU: Intersection over union; PET: Performance-efficiency trade-off; GRR: Generalization-reliability rate; IT: Information technology; PIs: Principal investigators; ML: Machine learning; QA: Quality assurance; DPIA: Data protection impact assessment; IRB: Institution review board; AUPRC: Area under precision-recall curve.