ImageNet pre-training and two-step transfer learning in chromosome image classification.

Area of Science:

• Computational biology
• Medical imaging analysis
• Machine learning for genomics

Background:

• Chromosome image classification commonly uses ImageNet pre-training.
• The utility of intermediate domain transfer learning for chromosome classification, particularly with related staining techniques, is not well-established.

Purpose of the Study:

• To evaluate the effectiveness of two-step transfer learning for chromosome image classification using Q-band and G-band datasets.
• To compare performance gains across different architecture families and training approaches.

Main Methods:

• Implemented two-step transfer learning, fine-tuning classifiers on an intermediate domain before the final task.
• Evaluated performance on Q-band (BioImLab) and G-band (CIR) chromosome datasets, using each as an intermediate domain for the other.
• Tested 11 architecture families and three training approaches.

Main Results:

• Modern architectures (ConvNeXt, Swin Transformer, ViT, MobileNetV3) showed significant Macro-F1 gains (+0.8 to +3.3 percentage points) on Q-band classification with limited data quality.
• Traditional CNNs showed minimal or no improvement on Q-band.
• On the higher-quality G-band dataset, all architectures achieved near-saturation performance with minimal gains (+0.1 to +0.7 percentage points) from two-step transfer.
• Consistent improvements were observed across both transfer directions.

Conclusions:

• Two-step transfer learning enhances chromosome classification performance when domain similarity is high and target data quality is limited, particularly with modern architectures.
• ImageNet pre-training alone is adequate for high-quality chromosome datasets.
• Architecture selection is crucial for maximizing benefits from transfer learning in chromosome image analysis.