Pyramid-based self-supervised learning for histopathological image classification.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Supervised learning in medical imaging requires large labeled datasets, which are difficult to obtain for histopathology.
• Self-supervised learning (SSL) offers a solution by pre-training models on unlabeled data.

Purpose of the Study:

• To propose a novel self-supervised Pyramid-based Local Wavelet Transformer (PLWT) model for enhanced feature extraction in histopathology.
• To evaluate the effectiveness of PLWT in pre-training models for downstream tasks using unlabeled histopathological images.

Main Methods:

• Developed the PLWT model incorporating wavelet transforms to reduce information loss during feature extraction.
• Integrated a Local Squeeze-and-Excitation (Local SE) module with an inverse residual in the feedforward network to capture local image information.
• Pre-trained the model on a large dataset of unlabeled histopathology images using a self-supervised approach.

Main Results:

• PLWT demonstrated competitive performance compared to other SSL methods on histopathological image analysis.
• The transferability of visual representations learned by PLWT on histopathology images surpassed that of a supervised model trained on ImageNet.
• Wavelet-based downsampling significantly reduced information loss in feature transmission.

Conclusions:

• The proposed PLWT model effectively extracts rich local and global features from histopathology images using self-supervised learning.
• PLWT shows strong potential for improving medical image analysis by leveraging unlabeled data.
• Self-supervised pre-training with PLWT enhances the transferability of learned representations for histopathology tasks.