Integrating quantitative phase imaging with deep learning for enhanced cervical cancer detection

期刊论文
发表日期：

2025年9月29日

查看全文

Ubaid Dar,
Assif Assad,
Muzafar Rasool,
Farooq Hussain,
Shabir Ahmad,
Mandeep Singh

摘要

In this study, we report a comparative analysis of brightfield (2D) images, raw holograms, and numerically reconstructed phase (3D) and amplitude images of cervical cancer samples for binary classification. The phase information extracted from the holograms exhibits a critical role in improving model performance. Three self-configured convolutional neural network (CNN) models: CNN \({}_{1}\) , CNN \({}_{2}\) , and CNN \({}_{3}\) of increasing architectural complexity were used. All of these models perform significantly better; in particular, the CNN \({}_{2}\) model exhibits the highest classification accuracy on both holographic and phase data. The classification efficiency is substantiated by model performance metrics. Additionally, the significance of phase images is further validated by the k-fold cross-validation technique. Therefore, this study underscores the potential of integrating holographic imaging into workflows to develop reliable, scalable, and accurate systems for cervical cancer detection.