Ocampo-López-Escalera José, Ochoa-Díaz-López Héctor, Sánchez-Chino Xariss M, Irecta-Nájera César A, Tobar-Alas Saúl D, Rosete-Aguilar Martha
Departamento de Salud, El Colegio de la Frontera Sur, San Cristóbal de las Casas, Chiapas, México.
SECIHTI - Departamento de Salud, El Colegio de la Frontera Sur, Villahermosa, Tabasco, México.
Front Med Technol. 2025 Mar 31;7:1531817. doi: 10.3389/fmedt.2025.1531817. eCollection 2025.
Cervical cancer remains a significant health challenge around the globe, with particularly high prevalence in low- and middle-income countries. This disease is preventable and curable if detected in early stages, making regular screening critically important. Cervical cytology, the most widely used screening method, has proven highly effective in reducing cervical cancer incidence and mortality in high income countries. However, its effectiveness in low-resource settings has been limited, among other factors, by insufficient diagnostic infrastructure and a shortage of trained healthcare personnel.
This paper introduces the development of a low-cost microscopy platform designed to address these limitations by enabling automatic reading of cervical cytology slides. The system features a robotized microscope capable of slide scanning, autofocus, and digital image capture, while supporting the integration of artificial intelligence (AI) algorithms. All at a production cost below 500 USD. A dataset of nearly 2,000 images, captured with the custom-built microscope and covering seven distinct cervical cellular types relevant in cytologic analysis, was created. This dataset was then used to fine-tune and test several pre-trained models for classifying between images containing normal and abnormal cell subtypes.
Most of the tested models showed good performance for properly classifying images containing abnormal and normal cervical cells, with sensitivities above 90%. Among these models, MobileNet demonstrated the highest accuracy in detecting abnormal cell types, achieving sensitivities of 98.26% and 97.95%, specificities of 88.91% and 88.72%, and F-scores of 96.42% and 96.23% on the validation and test sets, respectively.
The results indicate that MobileNet might be a suitable model for real-world deployment on the low-cost platform, offering high precision and efficiency in classifying cervical cytology images. This system presents a first step towards a promising solution for improving cervical cancer screening in low-resource settings.
宫颈癌仍然是全球一项重大的健康挑战，在低收入和中等收入国家发病率尤其高。如果在早期阶段发现，这种疾病是可预防和可治愈的，因此定期筛查至关重要。宫颈细胞学检查是使用最广泛的筛查方法，已证明在高收入国家能有效降低宫颈癌的发病率和死亡率。然而，由于诊断基础设施不足和训练有素的医护人员短缺等因素，其在资源匮乏地区的有效性受到限制。
本文介绍了一种低成本显微镜平台的开发，旨在通过实现宫颈细胞学玻片的自动读取来解决这些限制。该系统具有一个能进行玻片扫描、自动聚焦和数字图像捕捉的自动化显微镜，同时支持人工智能（AI）算法的集成。而其生产成本低于500美元。利用定制显微镜创建了一个包含近2000张图像的数据集，涵盖了细胞学分析中七种不同的宫颈细胞类型。然后使用该数据集对几个预训练模型进行微调并测试，以对包含正常和异常细胞亚型的图像进行分类。
大多数测试模型在正确分类包含异常和正常宫颈细胞的图像方面表现良好，灵敏度高于90%。在这些模型中，MobileNet在检测异常细胞类型方面表现出最高的准确率，在验证集和测试集上的灵敏度分别达到98.26%和97.95%，特异性分别为88.91%和88.72%，F值分别为96.42%和96.23%。
结果表明，MobileNet可能是在低成本平台上进行实际部署的合适模型，在宫颈细胞学图像分类方面具有高精度和高效率。该系统朝着在资源匮乏地区改善宫颈癌筛查的有前景的解决方案迈出了第一步。
