On the classification of a small imbalanced cytogenetic image database

Area of Science:

• Genetics
• Bioinformatics
• Machine Learning

Background:

• Multiclass classification with high-dimensional, imbalanced data presents significant challenges.
• The curse-of-dimensionality and training bias towards majority classes hinder accurate pattern recognition.
• Diagnosing genetic abnormalities using fluorescence in situ hybridization (FISH) signals exemplifies this problem.

Purpose of the Study:

• To propose and evaluate methods for improving multiclass classification accuracy on small, imbalanced datasets.
• To address the specific challenges encountered in classifying genetic abnormalities from FISH signals.
• To compare the effectiveness of different data handling strategies.

Main Methods:

• Hierarchical decomposition of the classification task into simpler, balanced subproblems.
• Data balancing via up-sampling minority classes combined with dimensionality reduction.
• Implementation and evaluation using naive Bayesian classifiers and multilayer perceptron neural networks.

Main Results:

• Both proposed solutions effectively mitigated classification difficulties.
• Accuracy improvements were observed with both hierarchical decomposition and data balancing techniques.
• Experimental results indicate that addressing data smallness is more impactful than managing data imbalance.

Conclusions:

• Hierarchical decomposition and data balancing are viable strategies for high-dimensional, imbalanced classification.
• The size of the dataset plays a more crucial role in classification performance than class imbalance.
• These methods offer improved accuracy for genetic abnormality diagnosis using FISH data.