Evolving Error States: Failure-Aware Progressive Repair for Ultrasound Lesion Segmentation

Ziliang Wang; XuJiang Tang; Lu Yuting; Weixin Xu; Yongqiang Zhao; Ying Fu; Kehua Guo

cs.CVSep 16, 2026

Evolving Error States: Failure-Aware Progressive Repair for Ultrasound Lesion Segmentation

Authors: Ziliang Wang, XuJiang Tang, Lu Yuting, Weixin Xu, Yongqiang Zhao, Ying Fu, Kehua Guo

Organizations: Central South University · Yangtze University · Chongqing University · Peking University ·

Abstract

Reliability under sparse and heterogeneous failures remains a fundamental challenge for medical image segmentation. High average accuracy can conceal a small set of structurally distinct and clinically consequential errors. Existing post-hoc correction methods alleviate this problem, but typically estimate false-positive and false-negative corrections from the same fixed prediction. This ignores the dynamic evolution of error states and limits the correction of complex cases. Inspired by iterative error feedback in structured prediction, we propose Failure-Aware Progressive Repair (FAPR). FAPR represents the current segmentation mask as a dynamic failure state and models each repair operation as a state-transition operator. Each accepted correction forms a new prediction state for subsequent error diagnosis and repair, enabling later operations to adapt to preceding changes. Conditional routing selectively activates necessary state transitions, while failure replay exposes the model to rare error states. By keeping the base segmentor frozen, FAPR preserves its established segmentation capability while improving difficult cases. Across three public ultrasound lesion segmentation benchmarks, FAPR improves mean DSC by 1.52%. On the very-hard subsets of BUSI and TN3K, the average gain reaches 13.77%.

Figures & tables

Figure 1: Persistent hard-case tails on BUSI. (a) Per-sample DSC distributions with model averages and parameter counts. (b) Among cases below 0.70 DSC, 17 are shared by CausalBridgeNet, U-KAN, and MSAGHNet.

Figure 2: Overview of FAPR. The frozen base model provides the initial mask $M_{0}$ and a shared multi-scale feature pyramid. Three failure-specific experts sequentially propose complete candidate masks $Q_{k}$ , and the corresponding gates selectively overlay each candidate on the latest state to produce $M_{1}$ , $M_{2}$ , and $M_{3}$ . A lightweight risk-aware controller finally accepts $M_{3}$ or rolls the prediction back to $M_{0}$ .

Method	Params (M)	DSC	mIoU	Prec.	Rec.	DSC	mIoU	Prec.	Rec.	DSC	mIoU	Prec.	Rec.
		BUSI				BUSIS				TN3K
U-Net (Ronneberger et al., 2015)	31.04	80.91	72.16	81.84	84.77	91.18	84.89	93.02	90.88	80.48	71.04	82.26	84.53
AttUNet (Oktay et al., 2018)	34.88	76.62	68.09	79.72	78.44	91.04	84.65	93.04	90.66	77.80	67.86	74.94	87.04
SegRefiner (Wang et al., 2023)	124.96	81.34	72.46	83.35	83.71	92.63	86.80	93.33	92.76	81.91	73.23	85.87	84.38
CMUNet (Tang et al., 2023)	49.93	79.95	71.68	84.13	81.45	91.43	85.28	92.86	91.58	79.86	70.15	78.35	85.19
MSAGHNet (Zhu et al., 2025)	26.61	81.44	72.53	84.04	81.78	91.36	84.76	92.88	90.94	81.01	71.35	79.83	87.55

Table 1: Comparison on BUSI, BUSIS, and TN3K. Values are percentages and Params denotes the total number of parameters in millions. The FAPR results are averaged over three random seeds; tiny subscripts denote standard deviations.

Dataset	Method	Very hard	Hard	Easy
BUSI	Stage 1	17.04	42.62	89.18
	SegRefiner	17.22	42.54	89.32
	CausalBridgeNet	17.20	43.80	89.58
	FAPR	35.78 $\scriptscriptstyle\pm 0.4$	53.69 $\scriptscriptstyle\pm 1.0$	90.11 $\scriptscriptstyle\pm 0.0$
TN3K	Stage 1	25.97	42.85	89.60
	SegRefiner	25.93	42.82	89.68

Table 2: Failure-conditioned DSC. Columns denote very-hard (

<0.5

), hard (

<0.7

), and easy (

\geq 0.7

) subsets. Their sizes are BUSI: 9/22/107 and TN3K: 53/102/512.

Figure 3: Qualitative prediction-state progression on two BUSI test cases. True-positive, false-positive, and false-negative regions are shown in green, red, and blue, respectively. (a) The strict-miss route is activated. (b) The miss route is bypassed, so $M_{1}=M_{0}$ , while the FN-expansion and FP-suppression specialists perform the subsequent repair.

Configuration	Overall	Easy	Hard	Very hard
FAPR	83.78 $\scriptscriptstyle\pm 0.2$	90.11 $\scriptscriptstyle\pm 0.0$	53.69 $\scriptscriptstyle\pm 1.0$	35.78 $\scriptscriptstyle\pm 0.4$
w/o Gross-miss	83.28	90.03	50.49	30.40
w/o FN Expansion	82.48	89.77	47.00	21.58
w/o FP Suppression	82.31	89.84	45.67	21.66
w/o Failure Replay	82.38	89.90	45.83	18.95
w/o Controlled Corruption	82.40	89.89	45.97	19.51

Table 3: Specialist and training ablations on the BUSI test set. The complete FAPR row includes the final rollback controller.

Configuration	Total	Train.	DSC	mIoU	Hard	Very hard
	(M)	(M)			DSC	DSC
U-KAN-Large (GB16)	25.36	25.36	78.76	68.85	47.36	28.96
Shared Refiner $\times 3$	17.32	10.70	82.69	74.17	48.01	27.77
Parallel Expert Fusion	17.61	10.99	82.99	74.41	48.84	29.21
Static- $M_{0}$ Cascade	17.37	10.75	82.98	74.45	50.88	33.61
Reverse Repair Order	17.37	10.75	82.55	73.92	46.02	21.83

Table 4: Mechanism ablation study on the BUSI test set.

Dataset	Method	DSC	mIoU	Hard	Very hard
BUSI	U-Net Stage1	80.91	72.16	38.36	16.38
BUSI	U-Net + FAPR	82.86	74.71	46.47	30.37
TN3K	U-Net Stage1	80.48	71.04	42.38	21.83
TN3K	U-Net + FAPR	82.55	73.22	52.19	37.83

Table 5: Backbone-transfer results with a frozen U-Net Stage-1 model.

Medical Image Segmentation

Medical image segmentation delineates organs, lesions, and other structures in scans to support measurement, diagnosis, treatment planning, and research.

See all 788 papers

Risk-Routed Implicit Boundary Refinement for Robust Ultrasound Image Segmentation
Jul 23, 2026Jingguo Qu, Xinyang Han, Xiang Wang +8Ultrasound Image Segmentation Implicit Neural Representations
Department of Health Technology and Informatics, The Hong Kong Polytechnic University, Hong Kong, China · Centre for Smart Health and School of Nursing, The Hong Kong Polytechnic University, Hong Kong, China · Department of Imaging and Interventional Radiology, The Chinese University of Hong Kong, Hong Kong, China
When Fusion Fails: Corruption-Aware Rebalanced Fusion for Multi-Modal Medical Image Segmentation
Sep 9, 2026Yuchen Pei, Xiaoyu Hu, Yixiong Zou +5Cross-Modality Medical Image Segmentation Multi-Source Image Fusion
Central China Normal University Wuhan, China · Huazhong University of Science and Technology Wuhan, China · Guangzhou University of Chinese Medicine Guangzhou, China +2
Implicit Fuzzification via Bounded Noise Injection for Robust Medical Image Segmentation
Jun 3, 2026Bisheng Tang, Zhangfeng Ma, Chuchu Zhai +4Ultrasound Image Segmentation Medical Image Segmentation
School of Information Science and Engineering, Provincial Key Laboratory of Informational Service for Rural Area of Southwestern Hunan, Shaoyang University, Shaoyang 422000, China · Xinshao County People’s Hospital, Shaoyang 422900, China

Jul 23, 2026J/K move · Enter open · S save

Risk-Routed Implicit Boundary Refinement for Robust Ultrasound Image Segmentation

Jingguo Qu, Xinyang Han, Xiang Wang, Yuqi Yang+7

Department of Health Technology and Informatics, The Hong Kong Polytechnic University, Hong Kong, China · Centre for Smart Health and School of Nursing, The Hong Kong Polytechnic University, Hong Kong, China · Department of Imaging and Interventional Radiology, The Chinese University of Hong Kong, Hong Kong, China

Ultrasound Image Segmentation Implicit Neural Representations Medical Image Segmentation Boundary-Aware Segmentation

Open paper AskPDF arXiv

Accurate boundary delineation in medical ultrasound (US) images remains challenging because of speckle noise, low contrast, acoustic shadowing, and acquisition variability. We propose Risk-routed Implicit Boundary Refinement (RIBR), a compact framework that refines convolutional segmentation through spatially weighted implicit residuals. A coordinate-conditioned implicit branch uses regional features to predict logit corrections, and a learned gate controls their contribution using predictive uncertainty and local appearance. Geometry- and speckle-aware regularization provides contour supervision and discourages probability fluctuations away from boundaries. We compare RIBR with ten baselines on nine datasets covering lymph nodes, breast lesions, thyroid nodules, and prostate, including four external lymph-node and breast cohorts. Across three seeds, RIBR achieves the best overall dataset macro-average among the evaluated configurations, with

76.67{\pm}0.42\%

Dice and a 95th-percentile Hausdorff distance (HD95) of

29.55{\pm}0.55

pixels on the resized evaluation grid. It obtains the lowest HD95 on all nine datasets, with task-dependent Dice rankings. These results support controlled implicit residual refinement for boundary-sensitive US segmentation. Source code is available at https://github.com/jinggqu/ribr.

Abstract

Medical Image Segmentation

Explore similar work

Risk-Routed Implicit Boundary Refinement for Robust Ultrasound Image Segmentation

When Fusion Fails: Corruption-Aware Rebalanced Fusion for Multi-Modal Medical Image Segmentation

Implicit Fuzzification via Bounded Noise Injection for Robust Medical Image Segmentation