SpatioPrompt: Learning Spatial Attention and Dynamic Prompts for Few-Shot Medical Image Anomaly Detection

We study normal-only few-shot medical image anomaly detection, where only a small support set of normal images is available for adaptation and no real abnormal samples are used during training. While recent vision-language models (VLMs) such as MediCLIP show promise through synthetic anomaly generation and adapter mechanisms, their linear projection architectures struggle to capture fine-grained spatial features, and static learnable prompts lack adaptivity to heterogeneous lesion patterns. We introduce SpatioPrompt, a parameter-efficient VLM adaptation framework that addresses these limitations through two complementary innovations. First, we incorporate a spatial attention mechanism inspired by CBAM to explicitly model local region dependencies, enhancing lesion-focused representations while suppressing background interference. Second, we propose a FewShotEnhancedRWKV decoder that fuses GRU-style gating with RWKV temporal recurrence, enabling dynamic generation of image-conditioned prompts. Experiments across three medical imaging modalities demonstrate substantial improvements over MediCLIP: 6.6% gain on BrainMRI (99.9% vs. 93.3% AUROC at 8-shot) and 3.9% gain on BUSI (92.0% vs. 88.1% at 4-shot), with consistent improvements on CheXpert (73.8% vs. 70.7% at 16/32-shot). These results show that combining spatial attention with dynamic prompt generation can improve normal-only few-shot medical anomaly detection.