APEX-SAM

Figure 1. Overview of APEX-SAM. QAR builds a hierarchical expert bank with DINO/SigLIP dual-stream indexing. APM performs style alignment, semantic gating, orientation-aware geometric alignment, and anatomy-aware point sampling. HMF fuses SAM prompt branches through training-free feature-consensus weighting.

Abstract

Training-Free Cross-Domain Segmentation

Training-free cross-domain few-shot medical image segmentation aims to segment unseen anatomies without parameter updates, addressing the high cost of dense annotation and domain-specific fine-tuning in clinical practice. Existing support-driven prompting methods face three limitations: support exemplars are randomly selected without quality assurance, geometric alignment is poorly modeled, and multi-modal prompt capabilities remain underexploited.

We present APEX-SAM, a retrieval-augmented framework with three innovations. QAR builds a dual-stream DINO/SigLIP expert bank with diversity-aware selection to ensure support-query compatibility. APM performs style-aligned geometric matching and anatomy-guided point sampling from morphological priors. HMF fuses SAM branches through training-free feature-consensus weighting. Experiments on three cross-domain benchmarks confirm strong performance among training-free methods, with ablations validating each component's contribution.

Method

Three Components, One Inference Pipeline

APEX-SAM improves support selection, prompt construction, and prompt-branch fusion while keeping the full pipeline training-free.

QAR

Quality-Aware Expert Retrieval

Builds a hierarchical expert bank using structural and semantic descriptors, then selects compatible support exemplars with quality, coverage, and diversity constraints.

APM

Anatomy-Aware Prompt Mining

Aligns support and query anatomy under cross-modality appearance shifts, then samples positive and negative prompts from morphological priors.

HMF

Hybrid Multi-Modal Fusion

Runs multiple SAM prompt branches and combines them with reliability-weighted feature consensus, avoiding learned fusion layers or fine-tuning.

Benchmarks

Cross-Domain Medical Segmentation

The paper evaluates APEX-SAM on Abd-MRI, Abd-CT, and Card-MRI, covering abdominal organs and cardiac structures with substantial modality and anatomy shifts.

Dataset visualization for Abd-MRI, Abd-CT, and Card-MRI

Figure 2. Dataset visualization and DINO-feature t-SNE for Abd-MRI, Abd-CT, and Card-MRI.

Results

Quantitative Comparison

APEX-SAM achieves consistent gains over training-free and few-shot baselines across abdominal and cardiac segmentation tasks.

Table 1. Dice (%) on Abd-MRI and Abd-CT. Bold indicates the best result and underline indicates the second best result.

Method	Ref.	Abd-MRI					Abd-CT
		Liver	LK	RK	Spleen	Mean	Liver	LK	RK	Spleen	Mean
PANet	ICCV'19	39.24	26.47	37.35	26.79	32.46	40.29	30.61	26.66	30.21	31.94
SSL-ALP	TMI'22	70.74	55.49	67.43	58.39	63.01	71.38	34.48	32.32	51.67	47.46
RPT	MICCAI'23	49.22	42.45	47.14	48.84	46.91	65.87	40.07	35.97	51.22	48.28
PATNet	ECCV'22	57.01	50.23	53.01	51.63	52.97	75.94	46.62	42.68	63.94	57.29
IFA	CVPR'24	50.22	35.99	34.00	42.21	40.61	46.62	25.13	26.56	24.85	30.79
FAMNet	AAAI'25	73.01	57.28	74.68	58.21	65.79	73.57	57.79	61.89	65.78	64.75
MAUP	MICCAI'25	78.16	58.23	72.34	59.65	67.09	78.25	59.41	71.80	60.38	67.46
APEX-SAM (Ours)	MICCAI'26	95.10	96.88	96.01	95.23	95.81	93.47	90.28	91.83	92.06	91.91

Table 2. Dice (%) on Card-MRI.

Method	Ref.	LV-BP	LV-MYO	RV	Mean
PANet	ICCV'19	51.42	25.75	25.75	36.66
SSL-ALP	TMI'22	83.47	22.73	66.21	57.47
RPT	MICCAI'23	60.84	42.28	57.30	53.47
PATNet	ECCV'22	65.35	50.63	68.34	61.44
IFA	CVPR'24	50.43	31.32	30.74	37.50
FAMNet	AAAI'25	86.64	51.82	76.26	71.58
MAUP	MICCAI'25	88.36	52.74	78.29	73.13
APEX-SAM (Ours)	MICCAI'26	92.75	68.41	88.23	83.13

Qualitative

Visual Results and Failure Analysis

APEX-SAM produces stable cross-domain masks while remaining transparent about remaining challenges such as low-contrast boundaries, severe shape shifts, retrieval mismatch, and suboptimal prompts.

Figure 3. Successful cross-domain masks and representative failure modes.

Ablation

Component Analysis

Each component contributes a measurable improvement, and the thresholded memory update further strengthens the full pipeline.

Table 3. Ablation on core designs (Dice %).

Configuration	QAR	APM	HMF	Memory	Mean Dice
Prompt-only baseline	No	No	No	-	72.4
+ QAR	Yes	No	No	Fixed	80.2
+ QAR + APM	Yes	Yes	No	Fixed	86.3
+ QAR + APM + HMF	Yes	Yes	Yes	Fixed	91.8
Full (Ours)	Yes	Yes	Yes	Thresholded	95.81

Figure 4. Hyperparameter sensitivity on Abd-MRI.

Mean Dice and standard deviation, with selected values highlighted.

Citation

Proceedings metadata such as LNCS volume, page range, DOI, and final paper URL will be added after publication.

Mao, Z., Chen, B., Lei, Q., Tan, J., Sun, K.: APEX-SAM: Anatomy-Aware Prompting with Expert Retrieval for Training-Free Medical Image Segmentation. In: Medical Image Computing and Computer Assisted Intervention - MICCAI 2026. Lecture Notes in Computer Science. Springer, Cham (2026).

@inproceedings{mao2026apexsam,
  title     = {APEX-SAM: Anatomy-Aware Prompting with Expert Retrieval for Training-Free Medical Image Segmentation},
  author    = {Mao, Zhihao and Chen, Bangpu and Lei, Qi and Tan, Jiaqi and Sun, Kun},
  booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
  series    = {Lecture Notes in Computer Science},
  publisher = {Springer},
  address   = {Cham},
  year      = {2026}
}