Computer assisted intervention to improve diagnostic yield in capsule endoscopy post ingestion.

PURPOSE: Wireless capsule endoscopy (WCE) examinations frequently become non-diagnostic due to early visibility degradation caused by dirt, bubbles, or sustained mucosal occlusion. Existing computational approaches typically assess examination quality retrospectively, after completion of the study, limiting the opportunity for corrective intervention. We propose a feasibility-aware computer-assisted intervention (CAI) framework that reformulates WCE quality management as a sequential decision problem during acquisition.

METHODS

Early checkpoints within the first 10-25% of each examination summarize frame-level visibility impairment using four structured components: mean impairment severity, temporal instability, worst-case continuous collapse, and dominant degradation mechanism cues. These early observations are integrated through logistic normalization to estimate downstream non-diagnostic risk without access to future frames. The framework further incorporates feasibility constraints through estimation of remaining diagnostically relevant transit time and intervention onset latency, enabling anatomically and temporally consistent action recommendations.

RESULTS

On the clinically annotated GALAR technical subset, early visibility dynamics from the first 20% of examinations achieved non-trivial discrimination of downstream non-diagnostic outcomes (AUROC ≈  0.78). While worst-window collapse alone achieved higher discrimination (AUROC 0.89), the proposed CAI risk maintained stable performance while preserving mechanism identifiability and decision stability required for intervention selection.

CONCLUSION

Capsule endoscopy quality management can be formally reframed as a sequential, feasibility-aware CAI problem. The proposed framework enables early, mechanism-consistent, and anatomically feasible intervention decisions rather than purely retrospective quality assessment.