Merging nucleus datasets by correlation-based cross-training.

Area of Science:

• Computational Biology
• Medical Imaging
• Machine Learning

Background:

• Fine-grained nucleus classification is hindered by high inter-class similarity and intra-class variability, necessitating large labeled datasets.
• Creating large-scale, consistently labeled nucleus datasets is difficult due to domain expertise requirements and existing dataset inconsistencies.
• Current methods train models separately on each dataset, limiting overall classification performance.

Purpose of the Study:

• To develop a unified framework for nucleus classification that leverages multiple, inconsistently labeled datasets.
• To address the challenge of missing labels across different nucleus classification datasets.
• To improve the performance of nucleus classification models by utilizing all available annotated data.

Main Methods:

• Formulated nucleus classification as a multi-label problem with missing labels, merging all available datasets.
• Devised a base classification module trained on all data with sparse supervision from ground-truth labels.
• Implemented a label correlation module to exploit relationships between different label sets and utilized cross-training with pseudo-labels.
• Incorporated data without ground-truth labels by generating pseudo-labels.

Main Results:

• The proposed framework demonstrated substantial performance improvements over state-of-the-art methods.
• Successfully unified and utilized multiple, inconsistently labeled nucleus classification datasets.
• Enabled training on data with missing labels and data lacking any ground-truth labels.

Conclusions:

• The multi-label, missing-label formulation provides an effective unified framework for nucleus classification.
• The proposed method overcomes limitations of inconsistent labeling criteria across datasets.
• This approach significantly enhances nucleus classification accuracy by maximizing data utilization.