Learning binary and sparse permutation-invariant representations for fast and memory efficient whole slide image...

Area of Science:

• Digital Pathology
• Computational Imaging
• Machine Learning for Medical Applications

Background:

• Whole slide images (WSIs) present significant challenges for computational analysis due to their gigapixel size.
• Current patch processing and multi-Instance Learning (MIL) methods demand high GPU memory during end-to-end training.
• There is a critical need for compact, sparse, or binary representations for efficient retrieval in large medical archives.

Purpose of the Study:

• To develop a novel framework for learning compact whole slide image (WSI) representations.
• To enhance memory and computational efficiency during WSI analysis and retrieval.
• To enable real-time image retrieval within large medical archives.

Main Methods:

• Utilized deep conditional generative modeling and Fisher Vector Theory for learning WSI representations.
• Employed instance-based training for improved memory and computational efficiency.
• Introduced gradient sparsity and gradient quantization losses to learn sparse (Conditioned Sparse Fisher Vector - C-Deep-SFV) and binary (Conditioned Binary Fisher Vector - C-Deep-BFV) permutation-invariant representations.

Main Results:

• The proposed C-Deep-SFV and C-Deep-BFV methods demonstrated superior retrieval accuracy and speed compared to Yottixel and GMM-based Fisher Vector on TCGA and LKS datasets.
• Achieved competitive performance against state-of-the-art methods for WSI classification on lung cancer data from TCGA and the LKS dataset.
• The instance-based training significantly reduced GPU memory consumption.

Conclusions:

• The novel framework effectively generates compact and efficient WSI representations suitable for large-scale search and classification.
• The proposed gradient sparsity and quantization losses are key to achieving sparse and binary representations for improved retrieval.
• This work addresses critical challenges in digital pathology, paving the way for more efficient medical image analysis systems.