Low Dimensional Representation of Fisher Vectors for Microscopy Image Classification

Area of Science:

• Biomedical imaging
• Computer vision
• Machine learning

Background:

• Accurate microscopy image classification is crucial for biomedical applications like cancer subtype identification and protein localization.
• Effective classification relies on representative and discriminative image feature descriptors.

Purpose of the Study:

• To develop a novel feature representation algorithm for automated microscopy image classification.
• To enhance the discriminative capability and reduce the dimensionality of image descriptors.

Main Methods:

• Incorporation of Fisher vector (FV) encoding with diverse local features (handcrafted and learned).
• Development of a separation-guided dimension reduction technique to optimize descriptor dimensionality and discriminative power.

Main Results:

• The proposed low-dimensional FV representation demonstrated consistent performance improvements.
• The method outperformed existing state-of-the-art techniques and common dimension reduction approaches.
• Evaluated on four diverse public microscopy datasets: UCSB breast cancer, MICCAI 2015 CBTC, IICBU malignant lymphoma, and RNAi.

Conclusions:

• The proposed method offers an advantageous approach for automated microscopy image classification.
• The enhanced feature representation and dimension reduction effectively improve classification accuracy in biomedical imaging.