ABSTRACT:

Cervical cancer remains a major global health concern, particularly in low-resource settings, where 88% of related deaths occur. To enhance early detection and improve screening accessibility, this study proposes a region-based approach for detecting cervical lesions in microscopic images. A low-cost, smartphone-based microscopy device is utilized for autonomous image acquisition and analysis of liquid-based cytology samples. Deep learning models for object detection were optimized and evaluated to determine the most effective architecture for lesion identification. Additionally, transfer learning from conventional cytology was explored to enhance detection robustness and mitigate challenges associated with the limited availability of mobile-acquired annotated datasets. A benchmark analysis on the SIPAKMED dataset yielded a test mean average precision (mAP) of 0.37798 and an average recall (AR) of 0.63651. Despite the need for further refinements before clinical deployment, the proposed method demonstrates promising results (cross-validation mAP of 0.20315, AR of 0.46572, and an analysis time of 4 minutes per sample), marking progress toward a cost-effective, automated cervical lesion detection framework.
Keywords: Cervical cancer, Early detection, Low-resource settings, Smartphone-based microscopy, Liquid-based cytology, Deep learning, Object detection, Region-based approach, Lesion detection, Transfer learning

INTRODUCTION:

Cervical cancer is one of the leading causes of cancer-related deaths among women, with the majority of these deaths occurring in low-resource settings. According to the World Health Organization, 88% of cervical cancer-related deaths are concentrated in these regions, where access to regular screening and medical care is limited. Early detection through cervical screening is critical for reducing mortality rates, but current methods often face significant challenges such as limited accessibility, high costs, and a shortage of trained cytologists.
Traditional cervical screening methods, such as pap smears and visual inspection, require specialized expertise and can be time-consuming. With the advancement of medical technology, automated methods using machine learning and deep learning techniques have gained attention for their potential to assist in early detection. However, current automated systems are often resource-intensive, requiring expensive equipment and infrastructure, making them less feasible in low-resource settings.
This study aims to address these challenges by proposing a region-based deep learning approach for the automated detection of cervical lesions in microscopic images of liquid-based cytology (LBC) samples. A low-cost, smartphone-based microscopy device is used to capture high-quality images of the samples, enabling autonomous image acquisition and analysis. This approach has the potential to make cervical cancer screening more accessible and affordable, particularly in regions where it is most needed.
Through the optimization and evaluation of deep learning models for object detection, this study seeks to identify the most effective architecture for cervical lesion identification. Additionally, transfer learning is explored to enhance detection accuracy and robustness, especially when dealing with the limited availability of annotated datasets from mobile-acquired images. The results of this work are expected to contribute toward the development of a cost-effective, automated screening tool that can improve early detection and save lives in underserved areas.

PROBLEM STATEMENT:

Cervical cancer screening is essential for early detection, but manual analysis of cytology samples is time-consuming and prone to errors. Existing automated methods struggle with lesion localization due to image variations and dataset limitations. To address this, a region-based deep learning approach is proposed to accurately detect and classify cervical lesions in microscopic images, improving efficiency and accessibility in cervical cancer screening.

OBJECTIVE:

The primary objective of this study is to develop and evaluate a region-based deep learning model for automated cervical lesion detection in microscopic images. The specific objectives are:
1. Develop a deep learning-based detection framework that can accurately locate and classify cervical lesions in LBC samples.
2. Compare different object detection models (Faster R-CNN, SSD, and RetinaNet) to determine the most effective meta-architecture and backbone combination for lesion detection.
3. Utilize transfer learning techniques to improve detection performance by leveraging pre-trained models and fine-tuning them for cervical cytology analysis.
4. Assess the generalizability of the model by applying the best-performing detection architecture, trained on the public SIPAKMED dataset, to mobile-acquired high-field-of-view (HFF) cytology images.
5. Optimize the model for practical deployment by balancing accuracy and computational efficiency to facilitate integration into a mobile or IoT-based cervical screening system.