Scholarlink

Background <p>Cervical cancer is a leading cause of cancer-related molarity among women in sub-Sahara Africa, where limited pathology capacity constrains timely diagnosis. Rwanda reports one of the highest cervical cancers incidence rates in the region, and there is a pressing need for scalable decision-support tools to augment histopathology services. This study evaluated deep learning models for automated binary classification of cervical histopathology images from a tertiary hospital in Kigali, Rwanda.</p> Methods <p>We conducted a retrospective, single-center study using 885 hematoxylin-and-eosin-stained cervical biopsy image patches acquired at the University Teaching Hospital of Kigali (CHUK) between 2018 and 2024. Regions of interest were annotated by board-certified pathologists and organized into normal versus abnormal (precancerous and malignant) categories. Three ImageNet-pretrained convolutional neural networks (ResNet50, EfficientNetB0, DenseNet121) were fine-tuned using patient-level, stratified splits (70% training, 20% validation, 10% test), class-weighted loss, and domain consistent data augmentation. Performance was assessed on the held-out test set using accuracy, sensitivity, specificity, F1-score, receiver operating characteristic area under the curve (ROC-AUC), Brier score, and calibration curves with 95% confidence intervals.</p> Results <p>EfficientNetB0 achieved the best overall performance, with test accuracy 0.99 (95% CI 0.97-1.00), sensitivity 0.98, specificity 1.00, F1-score 0.99, ROC-AUC 0.99, and the lowest Brier score (0.02). ResNet50 reached an accuracy of 0.91 (ROC-AUC 0.96), while DenseNet121 obtained an accuracy of 0.86 (ROC-AUC 0.96), with comparatively poorer calibration. EfficientNetB0 misclassified only one abnormal image as normal and produced no false positives on the test set.</p> Conclusions <p>Compact deep learning architectures, particularly EfficientNetB0, can deliver near-perfect discrimination between normal and abnormal cervical histopathology in a resource constrained Rwandan setting. These results support the feasibility of AI-assisted pathology as a triage tool but require external, multi-center and prospective validation before clinical deployment.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Image analysis for cervical cancer classification using deep learning techniques

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月23日

查看全文
  • Emmanuel Christian Nyabyenda,
  • Kizito Nkurikiyeyezu,
  • Muzungu Hirwa Sylvain,
  • Gashaija Absolomon,
  • Isaac Komezusenge,
  • Fauste Ndikumana,
  • Melissa Uwase,
  • Jean Damascene Hagenimana,
  • Piero Mazimpaka Irakiza,
  • Felix K. Rubuga,
  • Dieudonne Kayiranga,
  • Belson Rugwizangoga

摘要

Background

Cervical cancer is a leading cause of cancer-related molarity among women in sub-Sahara Africa, where limited pathology capacity constrains timely diagnosis. Rwanda reports one of the highest cervical cancers incidence rates in the region, and there is a pressing need for scalable decision-support tools to augment histopathology services. This study evaluated deep learning models for automated binary classification of cervical histopathology images from a tertiary hospital in Kigali, Rwanda.

Methods

We conducted a retrospective, single-center study using 885 hematoxylin-and-eosin-stained cervical biopsy image patches acquired at the University Teaching Hospital of Kigali (CHUK) between 2018 and 2024. Regions of interest were annotated by board-certified pathologists and organized into normal versus abnormal (precancerous and malignant) categories. Three ImageNet-pretrained convolutional neural networks (ResNet50, EfficientNetB0, DenseNet121) were fine-tuned using patient-level, stratified splits (70% training, 20% validation, 10% test), class-weighted loss, and domain consistent data augmentation. Performance was assessed on the held-out test set using accuracy, sensitivity, specificity, F1-score, receiver operating characteristic area under the curve (ROC-AUC), Brier score, and calibration curves with 95% confidence intervals.

Results

EfficientNetB0 achieved the best overall performance, with test accuracy 0.99 (95% CI 0.97-1.00), sensitivity 0.98, specificity 1.00, F1-score 0.99, ROC-AUC 0.99, and the lowest Brier score (0.02). ResNet50 reached an accuracy of 0.91 (ROC-AUC 0.96), while DenseNet121 obtained an accuracy of 0.86 (ROC-AUC 0.96), with comparatively poorer calibration. EfficientNetB0 misclassified only one abnormal image as normal and produced no false positives on the test set.

Conclusions

Compact deep learning architectures, particularly EfficientNetB0, can deliver near-perfect discrimination between normal and abnormal cervical histopathology in a resource constrained Rwandan setting. These results support the feasibility of AI-assisted pathology as a triage tool but require external, multi-center and prospective validation before clinical deployment.