Deep multiple-instance learning for abnormal cell detection in cervical histopathology images.

Area of Science:

• Pathology
• Medical Imaging
• Computer Science

Background:

• Cervical cancer diagnosis relies on manual histopathology, which is time-consuming and prone to errors.
• Automated analysis of cervical histopathology images is hindered by the high cost of whole-slide scanners, especially in low-resource settings.
• Developing affordable imaging systems and analysis algorithms is crucial for accessible cervical cancer screening.

Purpose of the Study:

• To assess the feasibility of a low-cost diagnostic system for H&E stained cervical tissue image analysis.
• To develop and evaluate automated image analysis algorithms for cervical precancer detection using smartphone-acquired images.
• To address the limitations of expensive whole-slide scanners in resource-limited regions.

Main Methods:

• Image acquisition using a smartphone attached to a light microscope for H&E stained cervical tissues.
• Formulation of the classification task as a deep multiple instance learning problem.
• Quantitative evaluation of four multiple instance learning algorithms with five different architectures and a sparse attention-based framework.

Main Results:

• A dataset of 1331 labeled cervical histopathology images was utilized.
• Multiple instance learning algorithms were trained and evaluated on images with varying optical magnifications.
• The developed sparse attention-based multiple instance learning framework achieved a maximum classification accuracy of 84.55% on the test set.

Conclusions:

• A low-cost, smartphone-based imaging system is feasible for cervical histopathology analysis.
• Deep multiple instance learning, particularly the sparse attention framework, shows promise for automated cervical precancer detection.
• This approach offers a potential solution for improving cervical cancer screening accessibility in resource-limited areas.