PaperlayerView exact revision on arXiv

Source-linked AI summary

A Dual Cross-Attention Framework for Colposcopic CIN Grading and Swede Score Prediction Using a New Multi-Center Dataset

Dania Khan, Nuzhat Aisha Shaikh, Asfina Hassan Juicy, Raiyun Kabir, S M Shahida, Taufiq Hasan

arXiv:2609.12827v1cs.CV

TL;DR

In LMICs, colposcopy is important for screening after abnormal primary tests, but access to trained specialists is limited and interpretation is subjective. The paper introduces a multi-center dataset and dual cross-attention models for CIN grading and Swede score prediction; it reports cross-component discrimination gains and a total-score MAE of 1.489.

  • Problem

    Colposcopy interpretation is subjective, trained specialists are scarce in LMICs, and representative multi-center colposcopy datasets are scarce.

    Source: §I · p. 1

  • Method

    The paper introduces a multi-center colposcopy dataset and dual cross-attention models for three-class CIN grading and multi-head prediction of the five Swede score components.

    Source: §I · pp. 1–2, F. Proposed Swede Score Prediction Model · p. 5

  • Results

    The discussion reports higher specificity and AUC-ROC across all five Swede components than baselines, while total Swede Score prediction has an MAE of 1.489.

    Source: B. Swede Score Component Prediction · p. 6, §VI · p. 8

  • Takeaways & Limitations

    The authors describe the framework as effective for resource-constrained clinical environments, while noting dataset size, imbalance, and heterogeneity challenges.

    Source: §VII · p. 8

    The study uses a small, severely imbalanced, heterogeneous multi-source dataset, and future work targets minority-class reliability.

    Source: §VII · p. 8

Abstract

from arXiv · show

Cervical cancer is a major global health challenge, with disease burden falling disproportionately on low- and middle-income countries (LMICs) due to a shortage of trained specialists and the subjective nature of colposcopy-based screening. To address this challenge, we propose a novel deep learning framework for the automated grading of Cervical Intraepithelial Neoplasia (CIN) and the prediction of clinical Swede scores. We also introduce the BUET Multi-Center Colposcopy Dataset, a novel, multi-center cohort designed and annotated for Swede score prediction and CIN grading. Our proposed dual-stream cross-attention architecture mimics the visual reasoning of an expert colposcopist by explicitly fusing paired multimodal cervigrams to evaluate comparative tissue responses. Furthermore, we introduce a custom composite loss function to address severe class imbalances and scoring inconsistencies across the five Swede score components. The proposed framework achieved 71.85% accuracy and an 86.23% AUC-ROC for three-class CIN grading, outperforming existing methods. For Swede score component prediction, the architecture achieved AUC-ROC values ranging from 75.7% to 88.4%, with the composite loss function yielding consistent F1-score improvements. Finally, the total predicted Swede Score, which ranges between 0 and 10, shows a Mean Absolute Error (MAE) of 1.489. The results show that the proposed method can pave the way towards developing AI-assisted colposcopy screening tools to support risk-based triage in resource-limited healthcare settings. The dataset and source code are publicly available(url: https://github.com/mHealthBuet/BUET-colposcopy)

I. INTRODUCTION

The introduction frames colposcopy as important in settings with limited screening access, but interpretation varies and representative annotated datasets are scarce. It presents a multi-center dataset and dual cross-attention framework for CIN grading and Swede score prediction, alongside a composite loss for imbalanced, cross-component scoring.

I. INTRODUCTION

  • Colposcopy is used as a main screening and biopsy-guidance tool after abnormal primary tests in LMICs, where trained specialists are scarce and interpretation varies even among trained colposcopists.The introduction also identifies a scarcity of well-annotated, multi-source datasets representative of LMIC populations.Source: §I · p. 1
  • The BUET Multi-Center Colposcopy Dataset was collected across multiple Dhaka healthcare facilities and is the country's first colposcopy dataset with transformation-zone and Swede-score annotations.The authors also curated the IARC Colposcopy Image Bank dataset.Source: §I · p. 1
  • The proposed dual-stream model performs three-class CIN grading by fusing paired cervigram features through swapped cross-attention, and a multi-head dual cross-attention network predicts individual Swede score components.Source: §I · p. 2
  • The proposed composite loss combines weighted focal loss, effective-number class weighting, a Huber-based total-score consistency term, and head correction to address class imbalance and cross-component scoring coherence.Source: §I · p. 2

A. CIN Classification

Prior CIN-classification work has progressed from static-frame classifiers to recurrent and attention-based multimodal methods, but single-institution evaluation and late feature concatenation remain common. Swede score prediction is less explored, with no peer-reviewed end-to-end deep-learning model for component prediction identified in the cited review.

A. CIN Classification

  • 96.13% accuracy was reported by C-RCNN across four classes, while Chen et al. reported 91.18% three-class accuracy with recurrent fusion versus 88.32% for the backbone alone.C-RCNN used 4,753 cervigrams from 679 cases; Chen et al. evaluated 6,002 cases and 18,006 images.Source: A. CIN Classification · p. 2
  • On independent test sets, a Swin Transformer/CNN model's AUC fell from 94% under five-fold cross-validation to 80–82%, illustrating an out-of-distribution generalization challenge.Source: A. CIN Classification · p. 2
  • Fig. 1 shows one case through normal-saline, green-filter, aceto-uptake, and iodine-uptake visualization modalities.Source: Figure 1 · p. 2
  • Most studies trained and validated on single-institution datasets, typically fused modalities by late concatenation, and did not combine class-aware reweighting with task-specific consistency constraints.Source: A. CIN Classification · p. 2

B. Swede Score Prediction

  • The cited Swede score literature includes AnnoCerv, a 527-image dataset from 100 clinical records, but identifies no peer-reviewed end-to-end deep-learning model for component prediction.Source: B. Swede Score Prediction · p. 2

A. IARC Colposcopy Image Bank

The paper describes the IARC image-bank subset, a new multi-center dataset from Dhaka hospitals, and the harmonized datasets used for CIN and Swede score modeling.

A. IARC Colposcopy Image Bank

  • The IARC subset contains 756 images from 189 case reports, with clinical metadata, structured Swede scores, four visualization states, and an imbalanced score distribution.The image bank supports development, evaluation, and validation of algorithms for early identification of cervical cancer and precancerous conditions.Source: A. IARC Colposcopy Image Bank · pp. 2–3

B. BUET Multi-Center Colposcopy Dataset

  • The BUET dataset includes 768 patient cases and 3,072 images collected across Dhaka hospitals, with transformation-zone and Swede score annotations.Selected CIN2 and CIN3+ cases received second-expert review; Table I summarizes the five Swede score characteristics, and the study reports ethics approvals and anonymization.Source: B. BUET Multi-Center Colposcopy Dataset · p. 3, Table I · p. 3

C. Combined Dataset

  • The combined datasets were harmonized for three-class CIN grading and five Swede components, yielding 957 cases and 3,828 images for Swede prediction.High Grade combines CIN2 and CIN3; missing component labels were excluded during Swede training, and Tables II and III present the resulting distributions.Source: C. Combined Dataset · p. 3, Table II · p. 3, Table III · p. 3

IV. METHODOLOGY

The methodology pairs colposcopic modalities in models for CIN classification and Swede score prediction, with expert annotation and modality-aware image preparation.

A. Overall Framework

  • The framework trains a three-class CIN classifier and a five-output Swede predictor whose component scores sum to a total from 0 to 10.The architecture uses paired multimodal cervigrams, parallel EfficientNet backbones, and swapped QKV cross-attention.Source: A. Overall Framework · p. 3, Figure 2 · p. 4, Figure 3 · p. 4

B. Data Annotation

  • Experienced colposcopists at participating hospitals retrospectively annotated individual Swede score components because routine reports often lack component-level scores..Source: B. Data Annotation · p. 3

1) Segmentation:

  • Manual freehand ROI segmentation delineated cervix boundaries and removed background and instrument artifacts across acetic acid, iodine, and green-filtered images..Source: 1) Segmentation · p. 3

2) Augmentation:

  • CIN image pairs received identical geometric augmentation and resizing, while Swede training stacked three modalities into nine channels for shared geometric augmentation before splitting them into inputs.CIN augmentation avoided photometric distortion to preserve diagnostic color responses; Swede augmentation also used zoom and contrast jitter.Source: 2) Augmentation · pp. 3–4

3) ROI cropping:

The model crops cervical tissue for Swede prediction and uses reciprocal cross-attention to integrate paired modalities into a joint diagnostic representation.

3) ROI cropping:

  • Swede prediction uses deterministic grayscale thresholding at τ = 15 to crop non-informative black background tightly around cervical tissue..Source: 3) ROI cropping · p. 4

D. Proposed Architecture

  • The framework is designed to emulate colposcopists’ diagnostic workflow by extracting and integrating complementary features from two modalities..Source: D. Proposed Architecture · p. 4

1) Swapped Query-Key-Value Cross-Attention:

  • Swapped cross-attention computes each modality’s query against the other modality’s keys and values to model cross-modal diagnostic correspondence.The authors state this lets the network identify spatial regions where both modalities jointly indicate pathological change rather than relying on independent processing followed by late fusion.Source: 1) Swapped Query-Key-Value… · p. 4

2) Bidirectional Feature Fusion:

The CIN classifier processes paired acetic acid and iodine images, fuses attended features, and refines the representation with attention before classification.

2) Bidirectional Feature Fusion:

3) Convolutional Block Attention Module:

E. Proposed CIN Classification Model

  • The CIN model uses ImageNet-pretrained EfficientNet-B4 streams, fuses intermediate features, applies CBAM, and classifies pooled features into three diagnostic classes.Patient-level GroupShuffleSplit reserves 20% of patients for testing, with five-fold StratifiedGroupKFold cross-validation on the remaining 80%.Source: E. Proposed CIN Classification Model · p. 5

F. Proposed Swede Score Prediction Model

The model predicts five Swede score components from paired colposcopic modalities and sums them into a total score. Its training framework uses case-level stratification and component-specific weight freezing alongside a loss designed to address imbalance and aggregate-score inconsistency.

F. Proposed Swede Score Prediction Model

G. Novel Composite Loss Function

1) Weighted Focal Loss:

The weighted focal-loss design emphasizes difficult samples and weights classes using an effective-number formulation. A Huber term aligns the summed component predictions with the true total score.

1) Weighted Focal Loss:

  • Focal Loss with γ = 2.0 down-weights easily classified samples and amplifies gradients from hard, typically minority-class samples.Source: 1) Weighted Focal Loss · p. 5
  • Effective Number of Samples class weights account for sample redundancy; β is tuned per head to imbalance severity, and inverse effective-number weights are normalized to sum to the number of classes.Source: 1) Weighted Focal Loss · p. 5

2) Total Score Huber Loss:

  • A Huber loss with δ = 1.0 penalizes the gap between the summed predicted component scores and the true total score.It is quadratic for small gaps and linear for larger errors, limiting gradients and improving robustness to clinical outliers.Source: 2) Total Score Huber Loss · pp. 5–6

3) Head Correction Loss:

The head-correction loss redistributes aggregate score error into adjusted targets for individual component heads. It is combined with focal and Huber losses to impose aggregate coherence without overriding head-level feature learning.

3) Head Correction Loss:

  • The head-correction mechanism assigns normalized weights from each head's relative error and creates pseudo-targets based on each head's proportional contribution to total-score error.The pseudo-targets are clipped to preserve the valid {0,1,2} scoring range; ε is used for numerical stability.Source: 3) Head Correction Loss · p. 6

4) Combined Loss:

  • The combined loss adds focal loss to λ times the Huber and correction losses, with λ = 0.2.The small scale factor softly imposes aggregate coherence without overriding fine-grained feature learning in individual heads.Source: 4) Combined Loss · p. 6

V. EXPERIMENTAL RESULTS

The proposed model outperformed conventional CNN baselines on CIN classification and achieved statistically significant AUC-ROC gains over EfficientNetB0 for three Swede components. The reported results also include total-score error and improvements from the custom composite loss.

B. Swede Score Component Prediction

A. CIN Classification

B. Swede Score Component Prediction

  • 0.766 vs 0.679, 0.884 vs 0.792, and 0.757 vs 0.632 AUC-ROC for Aceto Uptake, Iodine Uptake, and Vessel Pattern, respectively, versus EfficientNetB0.The reported p-values were 0.0019, 0.0456, and 0.0063, respectively; the proposed model performed best on the majority of metrics for every component.Source: B. Swede Score Component Prediction · p. 6, Table VI · p. 7

C. Effect of the Custom Composite Loss

VI. DISCUSSION

The discussion reports conservative Swede-score errors and strong discrimination across all five components, alongside CIN predictions described as avoiding dangerous under-referrals. The conclusion presents the framework and multi-center dataset while identifying data imbalance and unseen-center validation as remaining challenges.

VI. DISCUSSION

  • The CIN model is described as maintaining balanced sensitivity and high specificity, with extreme severe-lesion-to-normal errors highly rare and dangerous under-referrals prevented.The authors attribute minor overlap to subtle morphological transitions between adjacent pathological stages.Source: §VI · p. 8
  • Across all five Swede score components, the model achieves higher specificity and AUC-ROC than baselines, with gains attributed to the composite loss.The discussion also reports substantially improved MAE and RMSE, and rare Score 0-to-Score 2 confusion, including zero confusion for Lesion Size.Source: §VI · p. 8

VII. CONCLUSION

  • The conclusion presents a multi-center dataset and dual cross-attention framework, while identifying severe sub-score imbalance and unseen-center generalization as future-work targets.Planned steps include targeted data collection, resampling and loss reweighting, and testing on healthcare centers excluded from pooled training.Source: §VII · p. 8
Loading 2609.12827v1…