Neutrosophic Information Fusion of Imaging and Clinical

Evidence for Multimodal Disease Screening

Ali Refaat1,* Anvor Sulymanov2

1 Souhag University, Egypt

2 Faculty of Engineering, Central Asian University, Uzbekistan

Received: August 02, 2025 Accepted: November 05, 2025 ⋆ Corresponding author

ABSTRACT

Modern screening rarely relies on a single source of evidence: a clinician weighs an imaging finding against

laboratory markers and patient history, and these sources routinely conflict. We propose a neutrosophic informationfusion

framework that represents each evidence channel as a neutrosophic triple (t, i, f )—the degree to which the

channel supports disease, the degree to which it is indeterminate (noisy, borderline, missing), and the degree to

which it argues against disease. Channels are combined with a conflict-aware neutrosophic fusion rule that routes

disagreement into the indeterminacy component instead of silently averaging it away. A decision is issued only

when fused indeterminacy falls below a referral threshold; otherwise the case is escalated for expert review. On

a synthetic two-modality screening cohort the framework attains 91.8% accuracy while flagging 12% of cases as

indeterminate, and it degrades gracefully when one modality is corrupted—losing only 2.3 accuracy points against

6.1 for score-level averaging. A cost-sensitive analysis shows the referral gate lowers expected clinical cost when a

missed positive is much more expensive than a review.

Keywords: Neutrosophic sets Information fusion Multimodal diagnosis Decision support Uncertainty Selective

classification Medical screening

1. INTRODUCTION

Diagnostic reasoning is a fusion problem. A radiologist’s

read of an image, a panel of blood markers, and a structured

symptom questionnaire each provide partial and imperfect

evidence, and the art of diagnosis lies in combining them.

Automated decision-support systems that fuse such channels

tend to use one of two strategies: feature-level concatenation

followed by a single classifier, or score-level averaging of

per-modality classifiers. Both share a weakness. When one

channel says “disease” and another says “healthy,” averaging

produces a confident-looking middling score that hides the

conflict, and concatenation lets a corrupted channel silently

drag the joint prediction.

The clinical reality is that indeterminacy—a borderline image,

a haemolysed sample, an incomplete history—is not the same

as evidence for health. A screening tool that cannot tell “the

evidence points to healthy” from “the evidence is unreadable”

is dangerous, because the two demand opposite responses:

reassurance in the first case, escalation in the second. Neutrosophic

logic gives us a vocabulary for exactly this distinction

by carrying an independent indeterminacy term alongside

support and opposition. This paper develops that idea into

a working screening pipeline and, crucially, uses the fused

indeterminacy as a gate: cases the system cannot resolve are

routed to a human rather than forced into a binary label. This

connects the neutrosophic framework to the machine-learning

literature on selective classification, where a model is allowed