AI-assisted diagnosis of cervical dysplasia from cervicography images

Sep 22, 2026

Uploaded by James

Cervicography using Visual Inspection with Acetic Acid (VIA) supports cervical cancer screening in low-resource settings, but it cannot reliably grade dysplasia severity. This study introduces a multi-task learning framework with ensemble aggregation to estimate lesion severity from cervicography images using color, surface texture, lesion position, and lesion area across quadrants. Training uses Swede Score annotations and validation relies on histopathology-confirmed images from IARC. StyleGAN-2 with ADA augments data to address scarcity, improving accuracy from 62% to 95.21% with high sensitivity and precision, while noting limitations affecting generalizability.

LanguageEnglish

CategoryHealthcare

FormatPDF

Pages34

Reading Time1 h 26 min

AccessFree to download

Introduction

Cervical cancer and dysplasia grading

VIA-based cervicography and its limitations

AI and deep learning for medical image diagnosis

Challenges in applying AI to VIA-based data

[www. nature.com/scientificreports](www. nature.com/scientificreports) OPEN AI-assisted diagnosis of cervical dysplasia from cervicography images Siti Nurmaini1,2􀀍, Muhammad Naufal Rachmatullah1,2, PatiyusAgustiansyah3, Rizal Sanif3, Irawan Sastradinata3, Elly Matul Imah4, Annisa Darmawahyuni1,2, Bambang Tutuko1,2, Akhiar Wista Arum1,2, Anggun Islami1,2, Firdaus Firdaus1,2, Ade Iriani Sapitri1,2, Aini Nabilah1,2, Radiyati Umi Partan5 & Rizqi Ayunda Pratama6 Cervicography using Visual Inspection with Acetic Acid (VIA) is widely adopted for cervical cancer screening in low-resource settings. Although effective for identifying visible lesions, VIA cannot determine the severity of dysplasia, limiting its diagnostic utility. This study proposes a multi-task learning framework combined with an ensemble mechanism to estimate lesion severity directly from cervicography images. Four clinically relevant features—color, surface texture, lesion position, and lesion area across quadrants—learned through deep learning model and aggregated using ensemble method resulting a severity decision. Five datasets were used across training, testing, and prediction stages; Swede Score annotations supported model training, while histopathology-confirmed images from IARC were used for validation to ensure reliable ground truth. To address data scarcity, StyleGAN-2 with adaptive discriminator augmentation (ADA) was employed to generate synthetic images for augmentation. Initial multi-task learning experiments achieved 62% accuracy. After applying GAN-based augmentation and ensemble learning, performance improved substantially, reaching 95.21% accuracy, 95.08% sensitivity, and 81.25% precision for mild cases, and above 95% across all metrics for severe cases. Despite these promising results, the study has several limitations, including dataset imbalance, reliance on synthetic GAN-generated images, variability in imaging modalities across datasets, and a relatively small test set. These factors may affect the model’s generalizability in broader clinical applications. Overall, the findings highlight the potential of the proposed approach to enhance the diagnostic value of VIA-based cervicography, while underscoring the need for larger datasets and more extensive clinical validation. Keywords Dysplasia, Cervical precancer, Ensemble, Cervicography, VIA, SDGs Cervical cancer develops primarily from persistent infection with high-risk Human Papillomavirus (HPV), progressing through precancerous stages of dysplasia classified histopathologically as Cervical Intraepithelial Neoplasia (CIN 1, CIN 2, and CIN 3) before becoming invasive malignancy1–3. In clinical practice, distinguishing normal/CIN 1 from CIN2 + (CIN2–3) is essential because CIN 1 is usually monitored, whereas CIN2 + requires immediate treatment3,4. Colposcopy remains the reference diagnostic procedure for evaluating abnormal cervical findings and guiding biopsy1,2. Most prior work depends on digital colposcopy, which remains inaccessible in many rural and resource-limited regions2,4. In low-resource settings—such as Indonesia—the World Health Organization (WHO) recommends Visual Inspection with Acetic Acid (VIA) as a feasible alternative screening strategy5,6. VIAbased cervicography enables photographic documentation ofaceto-white changes but provides substantially less anatomic detail than colposcopy, making it inadequate for reliable dysplasia severity grading4,7,8. These inherent limitations highlight the need for improved diagnostic approaches to support VIA-based programs9. Artificial intelligence (AI) with deep learning (DL) methods have significantly advanced automated medical image interpretation across various specialties10–16. Among these advances, the You Only Look Once (YOLO) 1Intelligent System Research Group, Universitas Sriwijaya, Palembang, Indonesia. 2Artificial IntelligenceMedical Center of Excellence, Universitas Sriwijaya, Palembang, Indonesia. 3Department of

FAQs

What problem does the study target in VIA-based cervicography?

VIA-based cervicography can document aceto-white changes but cannot determine dysplasia severity, limiting its diagnostic usefulness.

How does the proposed model estimate lesion severity?

It uses a multi-task deep learning framework to learn features (color, surface texture, lesion position, and lesion area across quadrants) and combines predictions via an ensemble mechanism.

Why are synthetic images used and how do they affect performance?

StyleGAN-2 with adaptive discriminator augmentation generates synthetic images to reduce data scarcity and class imbalance, leading to substantial performance gains after augmentation and ensemble learning.

Sep 22, 2026

Uploaded by James