Scholarlink

<p>Unsupervised anomaly detection (UAD) is attractive for chest X-ray analysis because it can identify pathological deviations without requiring exhaustive abnormal annotations. Recent EHR-conditioned diffusion-based methods have shown strong performance by incorporating clinical information into the denoising process. However, two important limitations remain. First, existing methods usually depend on relatively heavy model configurations and prolonged training schedules, resulting in substantial computational cost and limiting practical deployment in resource-constrained medical environments. Second, they typically derive anomaly scores only from the final reconstruction error, discarding rich diagnostic evidence contained in intermediate denoising steps, dual-head consistency, and complementary residual patterns. To address these limitations, we propose <b>Timestep-conditioned Attention for Multi-dimensional Evidence (TAME)</b>, a unified framework with two complementary components. The first component, <b>Timestep-Conditioned Channel Attention (TCCA)</b>, is a lightweight architectural module that dynamically reallocates channel importance according to the current diffusion timestep and demographic/EHR conditions, enabling efficient training of a compact 96-channel model. The second component, <b>Multi-Dimensional Anomaly Evidence Fusion (MDAEF)</b>, is an inference-time scoring framework that aggregates anomaly evidence across five complementary dimensions: multiple timesteps, multiple scoring functions, multiple checkpoints, multiple calibration strategies, and multiple detection paradigms through a hybrid rank-based fusion strategy. Extensive experiments on CheXpert and MIMIC-CXR show that this unified design consistently delivers more accurate and resource-efficient medical anomaly detection, confirming the complementary benefits of TCCA and MDAEF.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Timestep-conditioned Attention and Multi-dimensional Evidence framework for efficient multimodal chest X-ray anomaly detection

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月8日

查看全文
  • Xueyu Kang,
  • Qiulan Liu,
  • Hailing Feng,
  • Qi Zhang,
  • Dongxia Lv,
  • Jun Wang

摘要

Unsupervised anomaly detection (UAD) is attractive for chest X-ray analysis because it can identify pathological deviations without requiring exhaustive abnormal annotations. Recent EHR-conditioned diffusion-based methods have shown strong performance by incorporating clinical information into the denoising process. However, two important limitations remain. First, existing methods usually depend on relatively heavy model configurations and prolonged training schedules, resulting in substantial computational cost and limiting practical deployment in resource-constrained medical environments. Second, they typically derive anomaly scores only from the final reconstruction error, discarding rich diagnostic evidence contained in intermediate denoising steps, dual-head consistency, and complementary residual patterns. To address these limitations, we propose Timestep-conditioned Attention for Multi-dimensional Evidence (TAME), a unified framework with two complementary components. The first component, Timestep-Conditioned Channel Attention (TCCA), is a lightweight architectural module that dynamically reallocates channel importance according to the current diffusion timestep and demographic/EHR conditions, enabling efficient training of a compact 96-channel model. The second component, Multi-Dimensional Anomaly Evidence Fusion (MDAEF), is an inference-time scoring framework that aggregates anomaly evidence across five complementary dimensions: multiple timesteps, multiple scoring functions, multiple checkpoints, multiple calibration strategies, and multiple detection paradigms through a hybrid rank-based fusion strategy. Extensive experiments on CheXpert and MIMIC-CXR show that this unified design consistently delivers more accurate and resource-efficient medical anomaly detection, confirming the complementary benefits of TCCA and MDAEF.