Semantic Iterative Reconstruction: One-Shot Universal Anomaly Detection

Ning Zhu

cs.CVMar 24, 2026

Semantic Iterative Reconstruction: One-Shot Universal Anomaly Detection

Authors: Ning Zhu

Abstract

Unsupervised medical anomaly detection is severely limited by the scarcity of normal training samples. Existing methods typically train dedicated models for each dataset or disease, requiring hundreds of normal images per task and lacking cross-modality generalization. We propose Semantic Iterative Reconstruction (SIR), a framework that enables a single universal model to detect anomalies across diverse medical domains using extremely few normal samples. SIR leverages a pretrained teacher encoder to extract multi-scale deep features and employs a compact up-then-down decoder with multi-loop iterative refinement to enforce robust normality priors in deep feature space. The framework adopts a one-shot universal design: a single model is trained by mixing exactly one normal sample from each of nine heterogeneous datasets, enabling effective anomaly detection on all corresponding test sets without task-specific retraining. Extensive experiments on nine medical benchmarks demonstrate that SIR achieves state-of-the-art under all four settings -- one-shot universal, full-shot universal, one-shot specialized, and full-shot specialized -- consistently outperforming previous methods. SIR offers an efficient and scalable solution for multi-domain clinical anomaly detection.

Figures & tables

Figure 1: Overview of the DSIR framework. The architecture consists of a frozen pre-trained teacher encoder $E$ and student decoder $D$ . During training, the teacher encoder processes one normal image from each of multiple source domains, extracting multi-scale deep features. The student decoder then performs iterative semantic reconstruction on the most compressed features through recurrent loops. At inference time, for any previously unseen target domain, the same single model receives only a test image $x$ , and the anomaly score and pixel-level anomaly map are finally obtained from the reconstruction discrepancies accumulated across all iterative loops.

Setting	Method	Venue	RSNA	OCT2017	APTOS	ISIC	BraTS	BR35H	LAG	VinCXR	Brain Tumor	Avg.
Full-shot specialized	RD4AD [ 12 ]	CVPR	74.7	97.3	90.2	68.3	74.8	99.6	67.5	57.8	93.2	80.4
	EDC [ 18 ]	TMI	85.9	99.7	95.8	89.3	66.6	99.8	58.1	60.7	96.9	83.6
	E2AD [ 35 ]	TMI	87.5	99.8	97.5	90.9	63.6	99.8	67.2	59.4	97.6	84.8
	SIR (Ours)	-	99.6	99.9	100.0	94.6	99.0	100.0	100.0	97.7	99.9	98.9
One-shot specialized	D24FAD [ 13 ]	LCLR	99.2	-	100.0	-	-	-	97.3	-	95.5	-
	INP-Former [ 28 ]	CVPR	78.5	98.5	91.1	76.1	90.8	99.4	75.2	65.9	81.4	84.1

Table 1: Image-level AUC (%) comparison on nine medical anomaly detection benchmarks. “Specialized” denotes per-dataset training, while “Universal” indicates a single model trained on the mixed source domains. Best results are in bold .

$L$	RSNA	OCT2017	APTOS	ISIC	BraTS	BR35H	LAG	VinCXR	Brain Tumor	Average
1	99.37	100.00	99.99	82.58	64.76	96.59	80.32	73.95	97.45	88.33
2	99.58	96.58	100.00	73.22	60.50	98.09	79.68	90.76	98.79	88.58
3	98.90	99.02	99.99	93.02	69.76	100.00	99.08	94.56	95.82	94.46
4	99.57	96.34	99.64	83.18	85.84	100.00	99.76	94.71	97.68	95.19
5	99.63	96.62	100.00	72.67	53.26	100.00	98.62	49.00	97.50	85.26

Table 2: Ablation on the number of iterative loops.

Metric	RSNA	OCT2017	APTOS	ISIC	BraTS	BR35H	LAG	VinCXR	Brain Tumor	Average
Loop 1 $f_{3}$	72.40	96.25	98.05	83.16	28.56	93.01	86.01	79.95	87.18	80.51
Loop 1 $\phi$	98.41	90.23	91.26	74.33	66.91	99.26	95.09	85.28	94.35	88.35
Loop 1 Fused	96.84	94.54	99.99	91.46	51.67	97.70	94.26	85.12	94.12	89.52
Loop 2 $f_{3}$	90.18	95.39	96.39	79.27	64.05	99.10	91.34	79.97	94.84	87.84
Loop 2 $\phi$	97.50	94.62	86.96	76.63	69.49	98.86	95.67	87.49	94.59	89.09
Loop 2 Fused	97.02	96.77	99.64	83.01	69.54	99.70	95.96	87.11	95.56	91.59

Table 3: Per-loop and fusion analysis (

L=3

) – averaged over 3 random seeds (final iteration).

Figure 2: Qualitative visualization of anomaly maps generated by DSIR under the one-shot universal detection setting. For normal samples, all four maps remain uniformly clean across loops. For abnormal samples, the anomaly regions become progressively sharper and more localized as the number of loops increases, with the final fused map providing the clearest boundary and highest contrast.

Train / Test	RSNA	VinCXR	APTOS	LAG	BR35H	Brain Tumor	BraTS
RSNA	99.62	65.63	-	-	-	-	-
VinCXR	99.54	90.08	-	-	-	-	-
APTOS	-	-	100.0	99.94	-	-	-
LAG	-	-	89.96	99.87	-	-	-
BR35H	-	-	-	-	100.00	98.41	40.80
Brain Tumor	-	-	-	-	100.00	98.38	62.52

Table 4: Intra-domain cross-dataset one-shot generalization.

Explore similar work

Dual Distillation for Few-Shot Anomaly Detection
Mar 2, 2026Le Dong, Qinzhong Tan, Chunlei Li +5Few-Shot Anomaly Detection Anomaly Detection
Xidian University, China · MedAI Technology (Wuxi) Co. Ltd., China · Technical University of Munich, Germany
A Feature Shuffling and Restoration Strategy for Universal Unsupervised Anomaly Detection
Mar 24, 2026Wei Luo, Haiming Yao, Zhenfeng Qiang +2Unsupervised Anomaly Detection Anomaly Detection
State Key Laboratory of Precision Measurement Technology and Instruments, Tsinghua University, Beijing, 100084, China · School of Medical Technology, Beijing Institute of Technology, Beijing, 100081, China
Diverse Normal Prototypes-Guided Contrastive Reconstruction for Medical Anomaly Detection
Aug 27, 2025Luhu Li, Bin Liu, Bowen Lin +3Prototype Learning Medical Image Anomaly Detection
School of Mathematics, Shandong University, Jinan 250100, China · Department of Interventional Medicine and Minimally Invasive Oncology, The Second Qilu Hospital of Shandong University, Jinan 250033, China · Department of Neurosurgery, The Second Qilu Hospital of Shandong University,2026 Jinan 250033, China

Mar 2, 2026J/K move · Enter open · S save

Dual Distillation for Few-Shot Anomaly Detection

Le Dong, Qinzhong Tan, Chunlei Li, Jingliang Hu+4

Xidian University, China · MedAI Technology (Wuxi) Co. Ltd., China · Technical University of Munich, Germany

Few-Shot Anomaly Detection Anomaly Detection Medical Image Anomaly Detection Knowledge Distillation

Open paper AskPDF arXiv

Anomaly detection is a critical task in computer vision with profound implications for medical imaging, where identifying pathologies early can directly impact patient outcomes. While recent unsupervised anomaly detection approaches show promise, they require substantial normal training data and struggle to generalize across anatomical contexts. We introduce D

^2

4FAD, a novel dual distillation framework for few-shot anomaly detection that identifies anomalies in previously unseen tasks using only a small number of normal reference images. Our approach leverages a pre-trained encoder as a teacher network to extract multi-scale features from both support and query images, while a student decoder learns to distill knowledge from the teacher on query images and self-distill on support images. We further propose a learn-to-weight mechanism that dynamically assesses the reference value of each support image conditioned on the query, optimizing anomaly detection performance. To evaluate our method, we curate a comprehensive benchmark dataset comprising 13,084 images across four organs, four imaging modalities, and five disease categories. Extensive experiments demonstrate that D

^2

4FAD significantly outperforms existing approaches, establishing a new state-of-the-art in few-shot medical anomaly detection. Code is available at https://github.com/ttttqz/D24FAD.

Abstract

Explore similar work

Dual Distillation for Few-Shot Anomaly Detection

A Feature Shuffling and Restoration Strategy for Universal Unsupervised Anomaly Detection

Diverse Normal Prototypes-Guided Contrastive Reconstruction for Medical Anomaly Detection