NoTAC: A Noise-Tolerance Automatic Cleaning Framework for Bone Marrow Karyotyping Data.

Area of Science:

• Bioinformatics
• Computational Biology
• Medical Imaging

Background:

• Deep neural networks (DNNs) are crucial for chromosome classification in karyotyping and disease diagnosis.
• Clinical chromosome datasets often suffer from label errors and outliers, hindering DNN performance and clinical use.

Purpose of the Study:

• To introduce NoTAC, a Noise-Tolerance Automatic Cleaning framework, designed to enhance DNN-based chromosome classification.
• To address and mitigate the impact of labeling errors and outliers in medical image datasets.

Main Methods:

• NoTAC employs a two-branch approach: KaryoCleanse for label noise detection and KaryoDrift for outlier identification.
• Label noise is detected using DNN self-confidence, latent label distribution estimation, and probability ranking.
• Outliers are identified and removed by scoring samples based on average K-nearest neighbor distances.

Main Results:

• NoTAC achieved a 93.99% accuracy on an R-band bone marrow chromosome dataset, a 6.25% relative improvement over baselines.
• The framework outperformed state-of-the-art noise-handling methods by 0.92%.
• Qualitative analysis revealed data issues impacting DNN predictions, offering insights into data quality's role.

Conclusions:

• NoTAC effectively enhances the performance and reliability of DNNs for practical medical datasets.
• The framework demonstrates potential for improving clinical karyotype diagnosis by ensuring data quality.