Determining Reliability in Surgical Visual Question Answering: Utilizing Question-Aligned Semantic Nearest Neighbor Entropy for Enhanced Safety
By
Luca Carlini
Dennis Pierantozzi
Mauro Orazio Drago
Chiara Lena
Cesare Hassan
Elena De Momi
Danail Stoyanov
Sophia Bano
Mobarak I. Hoque
July 17, 2026
Clinical Scorecard: Determining Reliability in Surgical Visual Question Answering: Utilizing Question-Aligned Semantic Nearest Neighbor Entropy for Enhanced Safety
At a Glance
Category Detail
Condition Surgical Visual Question Answering (VQA) Key Mechanisms Question-aligned semantic nearest neighbor entropy (QA-SNNE) for failure detection Target Population Surgeons and surgical teams in minimally invasive procedures Care Setting Preclinical evaluation of automatic failure detection in surgical VQA systems
Key Highlights
QA-SNNE combines answer consistency with question alignment for reliability assessment. The study introduces an out-of-template version of EndoVis18-VQA to evaluate model robustness. Semantic uncertainty methods are utilized for identifying unreliable outputs in VQA systems. Existing systems often fail to communicate uncertainty and may misinterpret clinical language. Guideline-Based Recommendations
Diagnosis
Management
Monitoring & Follow-up
Risks
Patient & Prescribing Data
Not applicable; focused on surgical VQA systems rather than direct patient treatment.
Emphasis on safety and reliability in surgical decision support systems.
Clinical Best Practices
Incorporate uncertainty detection mechanisms in surgical VQA systems. Utilize question-aligned scoring to enhance answer validity. Conduct preclinical evaluations to assess model performance under varied clinical language. Related Resources & Content
