Evaluating Multiple Multiview Fusion Strategies for Knee Fracture Detection in Paired Radiographs

Maximilian Nielsen, Finn Hinnerk Dieckhoff, Arne Ewald, Tobias Dust, René Werner

Keywords: Knee, Fracture, Classification, Multiview

TL;DR: Miultiview fusion with a self-attention mechanism and LoRA fine-tuning achieves the strongest performance for automated knee fracture detection from paired radiographic views

Abstract:

Background: Rapid and accurate evaluation of knee trauma in the Emergency Department is critical. While radiographs are the standard initial assessment, subtle fractures often lack overt visual signs, necessitating computed tomography (CT) for confirmation.

Objective: Unlike prior studies that focus on all types of knee fractures, this work addresses the automatic detection of diagnostically challenging, non-displaced knee fractures and explicitly investigates different view fusion strategies.

Methods: We evaluate multiple multiview deep learning frameworks that leverage complementary information from paired Anterior--Posterior and lateral X-ray projections. To address data scarcity and anatomical complexity, we employ radiology-specific self-supervised pretraining (RAD-DINO) combined with parameter-efficient fine-tuning via Low-Rank Adaptation (LoRA). We systematically evaluate different fusion strategies on a dataset of CT-confirmed knee fracture cases.

Results: Despite the high diagnostic difficulty of the cohort, our best-performing model (Self-Attention Fusion) achieves an AUROC of 0.88.

Conclusion: These findings demonstrate that combining multiview information with domain-adapted pretraining enables robust fracture detection in ambiguous cases.

Primary Subject Area: Application: Radiology

Secondary Subject Area: Detection and Diagnosis

Registration Requirement: Yes

Visa & Travel: No

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 299