Super-resolution recurrent convolutional neural networks for learning with multi-resolution whole slide images

Area of Science:

• Digital Pathology
• Medical Imaging
• Computer Vision

Background:

• Whole slide imaging (WSI) is crucial in digital pathology, but super-resolution (SR) algorithms face challenges with large resolution differences.
• Traditional SR algorithms typically use single high- and low-resolution image pairs, limiting their application in complex biomedical imaging.

Purpose of the Study:

• To develop an effective SR method for WSI by leveraging multi-resolution image data.
• To address the challenges of training SR models with limited data and significant resolution gaps in WSI.

Main Methods:

• A recurrent convolutional neural network (RCNN) model was proposed to generate super-resolved images from multi-resolution WSI datasets.
• The RCNN architecture was designed to exploit intermediate resolutions for improved learning tractability.

Main Results:

• The proposed RCNN model demonstrated state-of-the-art performance on three WSI histopathology cancer datasets.
• Utilizing intermediate resolutions significantly improved the trainability of the SR learning problem and handled large resolution differences effectively.
• The method achieved high performance even without requiring a large training dataset.

Conclusions:

• The RCNN model offers a robust solution for super-resolution in whole slide imaging.
• Leveraging multi-resolution data is a key strategy for enhancing SR performance in digital pathology applications.
• This approach shows promise for improving diagnostic accuracy and efficiency in computational pathology.