Domain transformation using semi-supervised CycleGAN for improving performance of classifying thyroid tissue images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning for medical image classification, particularly thyroid tissue images, requires substantial data.
• Data scarcity and variations in acquisition conditions across institutions hinder the direct application of deep learning models.
• Reusing models trained on data from one institution at another necessitates unifying feature distributions.

Purpose of the Study:

• To develop a domain transformation technique for unifying feature distributions in medical images from different institutions.
• To enhance the classification performance of thyroid tissue images by enabling cross-institutional model reuse.
• To address challenges posed by imbalanced datasets in medical image classification.

Main Methods:

• Employed semi-supervised CycleGAN for domain transformation, aligning feature distributions between Institution T and Institution S.
• Enhanced CycleGAN by incorporating class-specific feature distributions for improved domain adaptation.
• Integrated methods for handling imbalanced data within the semi-supervised CycleGAN framework.

Main Results:

• Domain transformation significantly improved classification performance when training on Institution S data and testing on Institution T data.
• The proposed method successfully retained class-relevant features across domains.
• Focal loss demonstrated the most effective improvement in mean F1 score for addressing class imbalance.

Conclusions:

• The developed domain transformation method effectively bridges feature distribution gaps between medical imaging datasets from different institutions.
• The approach enhances thyroid tissue image classification accuracy and robustness.
• Addressing class imbalance further optimizes the performance of the domain transformation technique.