Perturb-and-Restore: Simulation-driven Structural Augmentation Framework for Imbalance Chromosomal Anomaly Detection
Authors: Yilan Zhang, Hanbiao Chen, Changchun Yang, Yuetan Chu, Siyuan Chen, Jing Wu, Jingdong Hu, Na Li, Junkai Su, Yuxuan Chen, Ao Xu, Xin Gao, Aihua Yin
Organizations: Computer Science Program, Computer, Electrical and Mathematical Sciences and Engineering Division, King Abdullah University of Science and Technology (KAUST), Thuwal 23955-6900, Saudi Arabia · Center of Excellence on Smart Health (KCSH) and Center of Excellence for Generative AI, KAUST, Thuwal 23955-6900, Saudi Arabia · Guangdong Provincial Maternal and Child Health Hospital, Guangzhou 511442, China · Smiltec, Suzhou 215125, China
Detecting structural chromosomal abnormalities is crucial for accurate diagnosis and management of genetic disorders. However, collecting sufficient structural abnormality data is extremely challenging and costly in clinical practice, and not all abnormal types can be readily collected. As a result, deep learning approaches face significant performance degradation due to the severe imbalance and scarcity of abnormal chromosome data. To address this challenge, we propose a Perturb-and-Restore (P&R), a simulation-driven structural augmentation framework that effectively alleviates data imbalance in chromosome anomaly detection. The P&R framework comprises two key components: (1) Structure Perturbation and Restoration Simulation, which generates synthetic abnormal chromosomes by perturbing chromosomal banding patterns of normal chromosomes followed by a restoration diffusion network that reconstructs continuous chromosome content and edges, thus eliminating reliance on rare abnormal samples; and (2) Energy-guided Adaptive Sampling, an energy score-based online selection strategy that dynamically prioritizes high-quality synthetic samples by referencing the energy distribution of real samples. To evaluate our method, we construct a comprehensive structural anomaly dataset consisting of over 260,000 chromosome images, including 4,242 abnormal samples spanning 24 categories. Experimental results demonstrate that the P&R framework achieves state-of-the-art (SOTA) performance, surpassing existing methods with an average improvement of 8.92% in sensitivity, 8.89% in precision, and 13.79% in F1-score across all categories.
Figures & tables
Figure 1: Thoughts of Perturb-and-Restore (P&R). P&R mitigates the impact of long-tailed chromosomal distributions by simulating diverse structural anomalies from normal chromosomes. Guided by prior knowledge, structural perturbations are introduced to specific regions and refined via diffusion-based restoration. An energy-guided sampling strategy then selects high-quality synthetic anomalies to enhance the training of robust anomaly detectors.
Figure 2: The flowchart of the Structure Perturbation and Restoration Simulation (SPRS) module: (a) Chromosome skeletons are extracted and segmented into rectangular patches. (b) Structural perturbation operations are applied to rearranged chromosomes using prior knowledge.(c) The rearranged chromosomes and their original counterparts are paired to train a diffusion-based restoration model that treats rearrangement as image degradation. (d) The trained restoration model refines these synthetic anomalies by restoring realistic edges and content, generating highly realistic abnormal samples.
Figure 3: Framework of the proposed Perturb-and-Restore (P&R) method. (a) The SPRS module generates synthetic abnormal chromosomes by simulating structural anomalies based on prior abnormality knowledge. (b) The EAS module performs dynamic online sampling of synthetic data during training. By referencing the energy distributions of real normal and abnormal samples, the model selects high-quality synthetic anomalies to enhance detection performance.
Category
Train
Validation (200 Normal)
Test (800 Normal)
Unbalance Ratio
Synthetic Abnormality
1
343:10227
57
170
29.81
7926
2
72:10217
12
33
141.90
8438
3
42:10276
7
20
244.67
8799
4
51:10275
8
25
201.47
9136
5
44:10273
7
22
233.48
9269
6
83:10271
14
40
123.75
8913
Table 1: Summary of the chromosome used in this study. Each category corresponds to one of the 24 chromosome types (1–22, X, and Y). Values in the train column represent the number of abnormal versus normal chromosomes ( Abnormal:Normal ). The validation and test sets each include 200 and 800 normal samples per category. The unbalance ratio quantifies the level of class imbalance ( Normal/Abnormal ), while the synthetic abnormality column lists the number of synthetic abnormal chromosomes
Methods
Acc ↑
Sen ↑
Spe ↑
Pre(ab) ↑
Pre(n) ↑
AUC ↑
F1 ↑
ResNet18
Baseline
95.94 (2.79)
30.13 (31.21)
99.22 (1.76)
65.99 (44.67)
96.37 (2.23)
79.36 (11.79)
33.92 (31.20)
Under-sampling
91.39 (7.17)
50.38 (22.34)
93.39 (7.41)
41.43 (26.68)
97.22 (2.19)
83.99 (9.46)
38.27 (18.94)
Over-sampling
96.11 (3.28)
53.63 (15.98)
98.12 (2.59)
71.17 (19.67)
97.57 (1.75)
91.44 (4.87)
58.75 (13.32)
Focal Loss
95.33 (4.91)
54.52 (16.50)
97.96 (3.81)
70.88 (14.92)
96.01 (7.86)
91.16 (5.06)
59.44 (11.46)
LDAM-DRW
95.85 (4.70)
54.97 (15.64)
97.56 (5.04)
68.63 (13.48)
97.77 (1.38)
91.18 (4.65)
59.59 (11.83)
Table 2: Performance comparison of different methods for chromosomal anomaly detection (mean(std) %). The best results and the second-best results are highlighted in Bold and in underline .
Figure 4: Radar chart of chromosomal anomaly detection on 24 categories (sensitivity, precision of abnormality, and F1 score).
Figure 5: Performance comparisons under different imbalance ratio levels.
SYN
SYN*
EAS
Acc ↑
Sen ↑
Spe ↑
Pre(ab) ↑
Pre(n) ↑
AUC ↑
F1 ↑
p1
p2
ResNet18
95.94 (2.79)
30.13 (31.21)
99.22 (1.76)
65.99 (44.67)
96.37 (2.23)
79.36 (11.79)
33.92 (31.20)
⋆
-
✓
95.91 (3.00)
30.65 (24.13)
95.34 (16.88)
63.75 (28.18)
96.81 (1.57)
84.42 (10.38)
37.77 (24.79)
⋆ 0.328
△
✓
✓
96.42 (3.19)
55.30 (16.25)
98.42 (2.49)
71.90 (12.30)
97.65 (1.97)
90.75 (4.37)
60.88 (12.05)
⋆ 4.24×10−5
△ 3.77×10−6
✓
97.38 (2.64)
69.14 (16.90)
98.58 (2.84)
74.85 (12.76)
98.54 (0.94)
95.94 (3.51)
70.91 (13.62)
⋆ 1.19×10−6
▽
✓
✓
97.23 (2.98)
70.06 (12.04)
98.44 (2.79)
78.24 (11.97)
98.48 (1.07)
95.00 (2.75)
73.03 (9.54)
⋆ 1.51×10−7
▽ 0.226
Table 3: Performance comparison with different components. SYN represents the abnormality simulation data before restoration, while SYN* denotes the synthetic data after restoration. p1 indicates the p-value of the paired t-test relative to the baseline, while p2 represents the p-value relative to the model without EAS.
Methods
Acc ↑
Pre(ab) ↑
Sen ↑
F1 ↑
AUC ↑
CE
95.94 (2.79)
65.99 (44.67)
30.13 (31.21)
33.92 (31.20)
79.36 (11.79)
CE+SYN*
97.38 (2.64) ↑ 1.44
74.85 (12.76) ↑ 8.86
69.14 (16.90) ↑ 39.01
70.91 (13.62) ↑ 36.99
95.94 (3.51) ↑ 16.58
ReAD
96.29 (3.09)
73.86 (15.71)
48.61 (18.75)
56.17 (15.21)
91.21 (4.37)
ReAD+SYN*
96.58 (2.67) ↑ 0.29
76.87 (15.31) ↑ 3.01
53.41 (18.11) ↑ 4.8
60.67 (14.60) ↑ 4.5
91.39 (4.79) ↑ 0.18
Energy_ood
96.04 (3.57)
68.69 (16.71)
54.58 (17.81)
58.94 (14.39)
91.72 (5.54)
Energy_ood+SYN*
97.47 (2.60) ↑ 1.43
73.77 (13.05) ↑ 5.08
71.09 (17.44) ↑ 16.51
71.64 (14.59) ↑ 12.7
95.12 (3.78) ↑ 3.4
Table 4: Performance improvement across different methods with and without SYN*
Acc ↑
Pre(ab) ↑
Sen ↑
F1 ↑
AUC ↑
Coefficient λ
0.1
97.23 (2.98)
78.24 (11.97)
70.06 (12.04)
73.03 (9.54)
95.00 (2.75)
1
97.26 (2.22)
74.67 (10.97)
69.87 (10.53)
71.89 (9.79)
94.09 (3.48)
10
97.61 (2.46)
83.48 (10.48)
65.12 (15.07)
72.30 (12.12)
92.08 (5.77)
Interval k
5
96.74 (4.37)
70.55 (13.28)
73.84 (13.28)
71.54 (10.68)
94.77 (3.56)
Table 5: Anomaly detection performance under varying loss weights ( λ ) and sampling intervals ( k ).
Restoration
KID ↓
Iter
PSNR ↑
SSIM ↑
vs. Normal
vs. Abnormal
0
-
-
0.0412
0
10K
29.94
0.9687
0.0324
0.0460
50K
30.51
0.9678
0.0305
0.0342
100K
31.66
0.9708
0.0159
0.0318
150K
33.68
0.9768
0.0144
0.0325
Table 6: Restoration and detection performance at different iterations of training restoration model. For KID, the value at Iter 0 represents the divergence between real abnormal and normal chromosome images before restoration training.
Acc ↑
Group 1
Group 2
p -value
Per-user Acc
52.29 (5.20)
53.29 (6.76)
0.779
Acc ↑
Real Images
Synthetic Images
p -value
Per-image Acc
62.57 (15.36)
43.00 (14.81)
4.86×10−9
Table 7: Results of user study
Figure 6: Visualization of Energy score distributions for normal, abnormal, and synthetic data. (Top) Evolution of energy distributions for Chr1 over training epochs, showing progressive separation between normal (blue) and abnormal (red and orange) samples. (Bottom) Final energy score distributions across representative chromosomes with varying imbalance ratios (i.r.), demonstrating that the proposed EAS module maintains clear separability between normal and abnormal classes even under extreme imbalance conditions (please zoom in to view finer details).
Chromosome analysis is a fundamental step in the diagnosis of genetic diseases, but the manual karyotyping workflow is time-consuming and heavily dependent on expert specialists, often requiring several days per patient. Although Deep Learning models have achieved high performance in chromosome detection, most proposed solutions remain restricted to research prototypes or lack graphical interfaces suitable for clinical use. In this work, we present Aycromo, an open-source desktop platform for AI-assisted cytogenetic analysis. Built on Electron and ONNX Runtime, the tool allows cytogeneticists to load pre-trained models, compare architectures through an integrated benchmarking module, and manually correct detections via an interactive annotation interface, all without command-line interaction. Preliminary experiments on metaphase images from the CRCN-NE dataset demonstrate that YOLOv11 achieves 99.40% mAP@50, while the platform reduces per-slide analysis to seconds
Jorge L. A. Lima, Filipe R. Cordeiro
Visual Computing Lab, Department of Computing, Universidade Federal Rural de Pernambuco (UFRPE), Brazil
Data augmentation (DA) has been proven to be an effective means for improving protein representation learning (PRL) by generating additional training samples. Although mainstream perturbation- and sampling-based augmentation methods can produce data containing sufficient variations, they carry the risk of disrupting the protein structure and function. Some crafted protein homology modeling tools can generate conformations, but reduce structural diversity. The above dilemmas lead us to a question: Can we restore the disrupted structure caused by DA operations, providing data with both the original structure and diverse variations? In this work, we first analyze and empirically reveal the structure defect and performance degradation issues of existing DA methods. Based on the findings, we propose a simple yet effective DA method, Manifold Restore Mixing (MRM), for protein representation learning. Specifically, inspired by manifold mixup, we mix the hidden representations of original and augmented protein data to generate new samples that restore structural information lost in DA while introducing diverse variations. Furthermore, we develop a sample difficulty scheduler that adjusts the beta distribution in mixup to provide models with progressively challenging mixed samples during training, which improves the final performance. Comprehensive experiments on various PRL backbones and downstream tasks demonstrate the effectiveness and generalization of our method. The complete code and weights will be released upon acceptance. We provide a implementation at https://github.com/KingGugu/MRM.
Yizhou Dang, Chuang Zhao, Lianbo Ma +3
Software College, Northeastern University Shenyang, China · Tianjin University Tianjin, China · School of Computer Science and Engineering, Northeastern University Shenyang, China +1
Anomaly detection in structured images is challenging in small-data settings where deep learning approaches are costly or impractical. Classical template matching is simple and interpretable but lacks robustness to geometric variations such as scale, rotation, and perspective. We propose a variational template matching framework that represents anomaly templates as a family of transformed instances and performs detection via normalized cross-correlation over this transformation space. To further improve robustness, we introduce a density-based statistical anomaly score derived from local intensity distributions using kernel density estimation (KDE). This produces a smooth representation that captures distributional concentration and tail behavior more robustly than histogram-based methods. The structural and statistical signals are integrated through a unified fusion formulation, enabling complementary modeling of geometric similarity and distributional deviation. Experiments on biological cell images demonstrate that the proposed method outperforms classical baselines and achieves competitive performance with ResNet-50 under a fully training-free setting, while providing explicit localization. The approach offers an efficient, interpretable, and practical solution for anomaly detection in structured image domains.