Generalized deep learning for histopathology image classification using supervised contrastive learning

Md Mamunur Rahaman; Ewan K A Millar; Erik Meijering

doi:10.1016/j.jare.2024.11.013

Generalized deep learning for histopathology image classification using supervised contrastive learning

J Adv Res. 2024 Nov 16:S2090-1232(24)00532-0. doi: 10.1016/j.jare.2024.11.013. Online ahead of print.

Authors

Md Mamunur Rahaman¹, Ewan K A Millar², Erik Meijering³

Affiliations

¹ School of Computer Science and Engineering, University of New South Wales, Sydney, NSW 2052, Australia. Electronic address: md_mamunur.rahaman@unsw.edu.au.
² Department of Anatomical Pathology, NSW Health Pathology, St. George Hospital, Sydney NSW 2217, Australia; St. George and Sutherland Clinical School, University of New South Wales, Sydney NSW 2052, Australia; Faculty of Medicine & Health Sciences, Western Sydney University, Sydney NSW 2560, Australia. Electronic address: ewan.millar@health.nsw.gov.au.
³ School of Computer Science and Engineering, University of New South Wales, Sydney, NSW 2052, Australia. Electronic address: erik.meijering@unsw.edu.au.

PMID: 39551131
DOI: 10.1016/j.jare.2024.11.013

Abstract

Introduction: Cancer is a leading cause of death worldwide, necessitating effective diagnostic tools for early detection and treatment. Histopathological image analysis is crucial for cancer diagnosis but is often hindered by human error and variability. This study introduces HistopathAI, a hybrid network designed for histopathology image classification, aimed at enhancing diagnostic precision and efficiency in clinical pathology.

Objectives: The primary goal of this study is to demonstrate that HistopathAI, leveraging supervised contrastive learning (SCL) and hybrid deep feature fusion (HDFF), can significantly improve the accuracy of histopathological image classification, including scenarios involving imbalanced datasets.

Methods: HistopathAI integrates features from EfficientNetB3 and ResNet50, using HDFF to provide a rich representation of histopathology images. The framework employs a sequential methodology, transitioning from feature learning to classifier learning, mirroring the essence of contrastive learning with the aim of producing superior feature representations. The model combines SCL for feature representation with cross-entropy (CE) loss for classification. We evaluated HistopathAI across seven publicly available datasets and one private dataset, covering various histopathology domains.

Results: HistopathAI achieved state-of-the-art classification accuracy across all datasets, demonstrating superior performance in both binary and multiclass classification tasks. Statistical testing confirmed that HistopathAI's performance is significantly better than baseline models, ensuring robust and reliable improvements.

Conclusion: HistopathAI offers a robust tool for histopathology image classification, enhancing diagnostic accuracy and supporting the transition to digital pathology. This framework has the potential to improve cancer diagnosis and patient outcomes, paving the way for broader clinical application. The code is available on https://github.com/Mamunur-20/HistopathAI.

Keywords: Cancer diagnosis; Contrastive learning; Feature representation; Histopathological image analysis; Hybrid deep feature fusion; Imbalanced data.