ABO-Med: Accelerated Bilevel Optimization for Few-Shot Medical Image Classification

Ruoxuan Shi; Sheng Yang; Zhengxing Su; Xiaoyang Hou; Yating Liu

cs.CVSep 27, 2026

ABO-Med: Accelerated Bilevel Optimization for Few-Shot Medical Image Classification

Authors: Ruoxuan Shi, Sheng Yang, Zhengxing Su, Xiaoyang Hou, Yating Liu

Organizations: School of Mathematics and Statistics, Lanzhou University · Department of Statistics, University of California, Riverside · Department of Oncology, Lanzhou University Second Hospital

Abstract

In recent years, bilevel optimization has been widely used in a variety of machine learning tasks. However, prior bilevel optimization algorithms generally require the computation of second-order information, which limits their practical scalability. Only recently has a first-order paradigm for bilevel optimization been established, attaining near-optimal theoretical guarantees for solving bilevel optimization problems. In this paper, we propose ABO-Med, a scalable instantiation of this paradigm for few-shot learning, by incorporating it into the model-agnostic meta-learning (MAML) framework and tailoring it to medical image classification. We also introduce Medical Adaptive RandomAugment (MedRAug), a modality-aware augmentation strategy designed for medical images. Theoretically, ABO-Med establishes the optimality of MAML-type meta-learning approaches. Empirically, ABO-Med outperforms prior baselines on several public medical datasets, with gains of 1.99% to 18.76%, while MedRAug further improves the average accuracy by 2.20% to 6.34%. Additional cross-domain experiments, augmentation ablation studies, backbone ablation studies, and training efficiency analysis further validate the effectiveness and efficiency of the proposed method.

Figures & tables

Figure 2: Overall framework of ABO-Med, including task sampling, DataAugment, fully first-order task adaptation, meta-gradient construction, and outer-loop meta-update.

Method

Pap-Smear

ISIC2018

BreakHis40X

1-shot

3-shot

5-shot

7-shot

10-shot

1-shot

3-shot

5-shot

7-shot

10-shot

1-shot

3-shot

5-shot

7-shot

10-shot

MAML

64.59

\pm

0.97

76.97

\pm

0.76

81.78

\pm

0.55

84.96

\pm

0.45

87.41

\pm

0.39

42.36

\pm

0.70

48.09

\pm

0.70

48.86

\pm

0.65

51.23

\pm

0.62

54.90

\pm

0.60

37.88

\pm

0.66

44.45

\pm

0.65

47.31

\pm

0.62

47.17

\pm

0.57

49.31

\pm

0.61

FOMAML

64.60

\pm

1.09

78.16

\pm

0.77

81.21

\pm

0.59

84.11

\pm

0.48

85.88

\pm

0.43

36.73

\pm

0.55

47.64

\pm

0.67

51.36

\pm

0.65

52.04

\pm

0.65

52.73

\pm

0.60

38.15

\pm

0.71

47.01

\pm

0.69

49.64

\pm

0.63

55.93

\pm

0.67

49.63

\pm

0.61

Reptile

55.08

\pm

0.88

68.51

\pm

0.70

71.32

\pm

0.67

71.44

\pm

0.64

71.01

\pm

0.64

40.77

\pm

0.68

46.85

\pm

0.68

48.93

\pm

0.63

49.02

\pm

0.60

47.88

\pm

0.57

40.86

\pm

0.74

40.86

\pm

0.74

46.59

\pm

0.68

46.47

\pm

0.65

46.50

\pm

0.68

ALFA

66.40

\pm

1.16

78.33

\pm

0.59

82.17

\pm

0.46

81.85

\pm

0.42

87.56

\pm

0.36

42.73

\pm

0.75

52.57

\pm

0.64

57.12

\pm

0.63

56.14

\pm

0.64

61.57

\pm

0.63

37.09

\pm

0.64

52.41

\pm

0.66

54.28

\pm

0.64

56.16

\pm

0.64

60.38

\pm

0.57

Tra-MAML

66.40

\pm

1.02

81.48

\pm

0.64

84.42

\pm

0.47

86.71

\pm

0.40

87.51

\pm

0.37

41.71

\pm

0.66

48.67

\pm

0.68

55.20

\pm

0.66

61.25

\pm

0.61

62.84

\pm

0.56

38.45

\pm

0.67

44.96

\pm

0.62

57.76

\pm

0.66

62.61

\pm

0.67

62.67

\pm

0.60

Table 1: Comparison of various methods on the Pap-Smear, ISIC2018, and BreakHis40X datasets across different few-shot settings. Results are reported as mean test accuracy (%) with 95% confidence intervals.

Dataset	Method	Shot					Avg
		1	3	5	7	10
Pap-Smear	MAML	0.0711	0.0745	0.0760	0.0773	0.0801	0.0758
	FOMAML	0.0674	0.0706	0.0726	0.0732	0.0738	0.0715
	Reptile	0.0098	0.0112	0.0112	0.0134	0.0147	0.0121
	ALFA	0.0716	0.0820	0.0865	0.0956	0.1010	0.0873
	Tra-MAML	0.0597	0.0644	0.0696	0.0717	0.0722	0.0675

Table 2: Comparison of training time (h) per 100 tasks between few-shot learning methods and ABO-Med on the Pap-Smear, ISIC2018, and BreakHis40X datasets.

Method	BreakHis40X $\rightarrow$ BCHI (Near-Domain)			Pap-Smear $\rightarrow$ BCHI (Far-Domain)
	3-shot	5-shot	10-shot	3-shot	5-shot	10-shot
MAML	53.45 $\pm$ 0.66	56.13 $\pm$ 0.63	58.27 $\pm$ 0.62	47.46 $\pm$ 0.75	48.53 $\pm$ 0.70	56.18 $\pm$ 0.62
FOMAML	57.80 $\pm$ 0.76	62.24 $\pm$ 0.62	62.89 $\pm$ 0.52	46.08 $\pm$ 0.78	50.14 $\pm$ 0.72	56.31 $\pm$ 0.61
Reptile	55.27 $\pm$ 0.70	55.69 $\pm$ 0.64	56.78 $\pm$ 0.62	54.47 $\pm$ 0.69	55.08 $\pm$ 0.66	56.52 $\pm$ 0.58
ALFA	63.67 $\pm$ 0.67	65.13 $\pm$ 0.58	68.67 $\pm$ 0.53	46.03 $\pm$ 0.77	45.12 $\pm$ 0.73	53.06 $\pm$ 0.64
Tra-MAML	54.67 $\pm$ 0.68	68.68 $\pm$ 0.56	68.62 $\pm$ 0.52	52.48 $\pm$ 0.70	56.07 $\pm$ 0.62	57.46 $\pm$ 0.56

Table 3: Cross-domain few-shot classification results on the BCHI dataset under near-domain and far-domain settings. Results are reported as mean test accuracy (%) with 95% confidence intervals.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Hyperparameter	Value
Outer learning rate, $\eta$	1
Inner learning rate for $z$ update, $\alpha$	0.005
Inner learning rate for $y$ update, $\tau$	0.05
Lagrangian multiplier, $\lambda$	1
Number of inner iterations, $K$	30
number of adaptation steps $K^{\prime}$	30

Appendix

Table 4: Experimental hyperparameter settings for ABO-Med.

Backbone	Dataset	3-shot	5-shot	10-shot
CNN4 (0.113M)	Pap-Smear	81.51 $\pm$ 0.53	86.78 $\pm$ 0.65	87.62 $\pm$ 0.61
	ISIC2018	53.56 $\pm$ 0.65	58.28 $\pm$ 0.52	63.67 $\pm$ 0.64
	BreakHis40X	55.37 $\pm$ 0.83	69.07 $\pm$ 0.79	76.17 $\pm$ 0.89
ResNetFeats10 (4.898M)	Pap-Smear	75.57 $\pm$ 0.68	80.69 $\pm$ 0.55	84.18 $\pm$ 0.52
	ISIC2018	52.81 $\pm$ 0.64	55.59 $\pm$ 0.61	61.68 $\pm$ 0.63
	BreakHis40X	53.46 $\pm$ 0.63	65.92 $\pm$ 0.57	73.65 $\pm$ 0.54

Appendix

Table 5: Ablation study of different feature backbones on the Pap-Smear, ISIC2018, and BreakHis40X datasets. The number of trainable parameters of the feature extractor is reported below each backbone name. Results are presented as mean test accuracy (%) with 95% confidence intervals, and the best results are shown in bold.

Figure 3: Examples of different data augmentation strategies on the Pap-Smear, ISIC2018, and BreakHis datasets.

Operation	Type	Rationale
Identity	Baseline operation	Keeps the original cell image unchanged and avoids unnecessary structural perturbation.
Rotate	Geometric transformation	Simulates cell orientation variation, since cytology images usually have no fixed directional semantics.
Scale	Geometric transformation	Simulates differences in cell size and microscopic magnification.
Brightness	Photometric transformation	Simulates illumination and exposure variation in microscopic imaging.
Contrast	Photometric transformation	Simulates contrast differences caused by staining and imaging conditions.
RGBShift	Color transformation	Applies mild channel-wise color perturbation to simulate staining and color response variation.

Appendix

Table 6: Augmentation operation space for Pap-smear cytology images.

Operation	Type	Rationale
Identity	Baseline operation	Preserves the original dermoscopic image and avoids over-augmentation.
Rotate	Geometric transformation	Simulates lesion orientation variation caused by camera angle or patient positioning.
Translate-X	Geometric transformation	Simulates horizontal lesion position variation within the image.
Translate-Y	Geometric transformation	Simulates vertical lesion position variation while preserving lesion appearance.
Brightness	Photometric transformation	Simulates illumination differences under different imaging conditions.
Contrast	Photometric transformation	Improves robustness to contrast variation between lesion and surrounding skin.

Appendix

Table 7: Augmentation operation space for ISIC dermoscopic images.

Operation	Type	Rationale
Identity	Baseline operation	Keeps the original image unchanged and prevents excessive augmentation.
Rotate	Geometric transformation	Simulates tissue orientation variation caused by slide preparation and scanning.
Translate-X	Geometric transformation	Simulates horizontal position shifts of tissue regions.
Translate-Y	Geometric transformation	Simulates vertical position shifts of tissue regions.
Shear-X	Geometric transformation	Introduces mild horizontal shape variation while preserving tissue morphology.
Shear-Y	Geometric transformation	Introduces mild vertical shape variation while maintaining global tissue structure.

Appendix

Table 8: Augmentation operation space for BreakHis pathological images.

Dataset	Method	1-shot	5-shot	10-shot
Pap-Smear	Baseline	81.51 $\pm$ 0.68	86.78 $\pm$ 0.65	87.62 $\pm$ 0.61
	Cutmix	71.37 $\pm$ 0.66	74.58 $\pm$ 0.65	80.11 $\pm$ 0.54
	Mixup	70.56 $\pm$ 0.62	72.39 $\pm$ 0.58	79.46 $\pm$ 0.51
	Cutout	73.67 $\pm$ 0.69	80.47 $\pm$ 0.64	82.56 $\pm$ 0.62
	MedRAug	83.11 $\pm$ 0.67	87.42 $\pm$ 0.65	90.45 $\pm$ 0.58
ISIC2018	Baseline	53.56 $\pm$ 0.65	58.28 $\pm$ 0.62	63.67 $\pm$ 0.64

Appendix

Table 9: Ablation study of baseline data augmentation strategies, Cutmix, Mixup, Cutout, and MedRAug, on the Pap-Smear, ISIC2018, and BreakHis40X datasets. The results are reported as mean test accuracy (%) with their 95% confidence intervals.

Explore similar work

Benchmarking Foundation and Large Language Models for Few-Shot Medical Image Segmentation
Jul 30, 2026Jinghong Liu, Yuchuan Deng, Fanping Liu +2Medical Imaging Foundation Models Few-Shot Medical Image Segmentation
Renmin University of China
Bi-Level Collaborative Learning for Few-Shot Scribble-Supervised Medical Image Segmentation
Jul 28, 2026Xiang-Xiang Su, Yufan Ye, Yihang Zheng +2Weakly Supervised Image Segmentation Semi-Supervised Medical Image Segmentation
Fuzhou University Fuzhou, China · Qingdao University Qingdao, China
Semi-Supervised Few-Shot Adaptation of Vision-Language Models
Mar 3, 2026Julio Silva-Rodríguez, Ender KonukogluVLM Adaptation Few-Shot Image Classification
Computer Vision Lab, ETH Zurich, Zurich, Switzerland

Jul 30, 2026J/K move · Enter open · S save

Benchmarking Foundation and Large Language Models for Few-Shot Medical Image Segmentation

Jinghong Liu, Yuchuan Deng, Fanping Liu, Meng Huang+1

Renmin University of China

Medical Imaging Foundation Models Few-Shot Medical Image Segmentation Medical Image Benchmarks

Open paper AskPDF arXiv

Few-shot medical image segmentation (FS-MIS) aims to segment novel regions of interest (ROIs) from a few annotated support examples. Despite rapid progress, existing FS-MIS solutions span diverse paradigms but are evaluated under inconsistent settings, leaving their relative effectiveness unclear. We introduce FAME, a unified benchmark for evaluating FS-MIS solutions, covering specialists, SAM-based methods, CLIP-based methods, and MLLM-based methods. FAME contains 14,958 test samples across 7 anatomical sites, 9 imaging modalities, and 14 ROI categories, and evaluates models under zero-shot and ten-shot settings with additional assessment of target-absence recognition and generalization under covariate and semantic shifts. Our evaluation reveals several findings. First, effective few-shot segmentation depends on how models exploit support examples: direct visual adaptation generally outperforms prompt-based strategies. Second, increasing support examples improves performance only when models can effectively utilize them. Third, semantic transfer remains substantially more challenging than imaging-domain adaptation, and strong localization ability does not necessarily imply reliable target-absence recognition. We hope FAME provides a comprehensive understanding of current FS-MIS solutions and facilitates the development of more effective and reliable few-shot medical segmentation methods.

Abstract

Explore similar work

Benchmarking Foundation and Large Language Models for Few-Shot Medical Image Segmentation

Bi-Level Collaborative Learning for Few-Shot Scribble-Supervised Medical Image Segmentation

Semi-Supervised Few-Shot Adaptation of Vision-Language Models