Diagnostic Performance of a Locally Deployed Vision Language Model for Bone Tumor Diagnosis Using Smartphone-Captured Images: Exploratory Retrospective Study

Background Vision language models (VLMs) show promise in medical imaging, yet their performance on high-noise smartphone-captured images—common in primary care referrals—remains untested. Furthermore, it remains controversial whether retrieval-augmented generation (RAG) using external expert guidelines actually improves diagnostic accuracy for rare bone tumors. Objective This study aims to evaluate the diagnostic efficacy of a locally deployed, open-source VLM (Qwen3-VL) on high-noise bone tumor images. Specifically, we investigated how clinical persona prompts and RAG integration were associated with diagnostic performance and observed error patterns during multimodal reasoning. Methods This retrospective study included 42 patients with biopsy-proven primary bone tumors and tumor-like lesions. To simulate real-world conditions, we captured the original DICOM (Digital Imaging and Communications in Medicine) images from a monitor using a handheld smartphone without stabilization, organically capturing ambient glare and Moiré patterns typical of real-world teleconsultations. Using a 2×2 factorial design, we compared the diagnostic performance of the base model versus the RAG-integrated model under 2 distinct system personas: “radiologist” and “orthopedic oncologist.” Primary outcomes were top-1 and top-3 diagnostic accuracy. We used the McNemar test for paired comparisons of diagnostic correctness before and after RAG integration, and conducted an exploratory analysis of AI hallucinations using model-generated reasoning traces. Results Without RAG, the top-3 accuracy showed no significant difference between the radiologist and orthopedic oncologist personas (15/42, 36% vs 14/42, 33%; P=.76). After RAG integration, top-1 accuracy in the radiologist persona decreased from 12 (29%, 95% CI 16%-45%) to 6 (14%, 95% CI 5%-29%; P=.03). In the orthopedic oncologist persona, no significant change was observed in top-1 or top-3 accuracy (P=.65 and P>.99, respectively). Exploratory review of model-generated reasoning traces identified several text-associated diagnostic error patterns, including demographic anchoring, trauma-related masking, and shifts toward rare diagnoses following RAG retrieval. Conclusions In this exploratory study, adding the evaluated RAG configuration did not improve diagnostic accuracy and was associated with text-related diagnostic errors under smartphone-captured degraded imaging conditions. The findings suggest that persona design may influence the robustness of VLM responses to retrieved information, but this observation requires validation in larger, multimodel studies. Future studies should evaluate whether targeted visual fine-tuning on representative real-world degraded medical images can improve robustness under such conditions.