Enhancing Cervical Cancer Screening in Low-Resource Settings: A Deep Learning-Based Decision Support System Using Colposcopic Images



Yarandi F ; Feizabad E ; Shirali E ; Ramhormozian S ; Haghi A ; Khoshdooni Farahani A ; Rahmani A ; Nikoosokhan A
Source: Tehran University Medical Journal Published:2026

Abstract

Background: Cervical cancer is a common malignancy among women, and timely detection of precancerous lesions plays a crucial role in reducing mortality. This study aimed to evaluate and compare the performance of various deep learning architectures in classifying colposcopy images into normal and abnormal groups, and to assess their potential as a supportive tool for patient triage. Methods: In this analytical comparative study, 768 colposcopic images from patients referred with abnormal screening results (Pap smear or high risk HPV) were examined. Based on histopathological findings, images were divided into normal (47%) and abnormal (53%) groups and allocated patient wise into training, validation, and test sets. Six deep learning architectures-MobileNetV3 Large, EfficientNetV2 S, DenseNet 121, SE ResNet 50, ResNet 50, and ConvNeXt Tiny were trained, and their performance was evaluated based on sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). Results: Among the evaluated models, the ConvNeXt-Tiny architecture achieved the best performance, with a sensitivity of 93.3% (95% CI: 87.6-96.9), specificity of 84.4% (95% CI: 76.8-90.4), PPV of 86.8% (95% CI: 80.2-91.9), and NPV of 92.0% (95% CI: 85.3-96.3). On the test set, this model correctly identified 125 out of 134 abnormal cases (9 false negatives) and correctly classified 103 out of 122 normal cases (19 false positives). Conclusion: The ConvNeXt-Tiny deep learning model can serve as an effective supportive tool for colposcopy image analysis. By reducing dependence on individual expertise and promoting standardized interpretation, this approach may improve the quality of patient triage, particularly in regions with limited access to specialists. Multicenter studies are required to confirm the generalizability of these findings. Copyright © 2026 Yarandi et al. Published by Tehran University of Medical Sciences. This work is licensed under a Creative Commons Attribution-Non-Commercial 4.0 International license (https://creativecommons.org/licenses/by-nc/4.0/). Non-commercial uses of the work are permitted, provided the original work is properly cited.