Deep Transfer Learning-Based Approach for Glucose Transporter-1 (GLUT1) Expression Assessment

Area of Science:

• Oncology
• Medical Imaging
• Computational Pathology

Background:

• Glucose transporter-1 (GLUT-1) expression is a key biomarker for tumour hypoxia in immunohistochemistry (IHC) images.
• Current visual assessment of GLUT-1 scores is subjective, leading to inter-pathologist variability in clinical practice.
• Accurate GLUT-1 scoring is crucial for predicting tumour hypoxia markers.

Purpose of the Study:

• To develop and evaluate an automated method for assessing GLUT-1 scores in IHC colorectal carcinoma images.
• To compare the performance of various deep transfer learning models for GLUT-1 scoring.
• To establish a more objective and reliable approach for tumour hypoxia assessment.

Main Methods:

• Leveraged deep transfer learning with six pre-trained convolutional neural network (CNN) architectures (AlexNet, VGG16, GoogleNet, ResNet50, DenseNet-201, ShuffleNet).
• CNNs were fine-tuned as classifiers or used as feature extractors with a support vector machine (SVM).
• Evaluated classification of GLUT-1 scores in IHC images using automated feature extraction and classification.

Main Results:

• The best performing model was an SVM classifier using fused deep features (Feat-Concat) from DenseNet-201, ResNet50, and GoogLeNet.
• This model achieved a high prediction accuracy of 98.86%, surpassing other tested classifiers.
• Off-the-shelf feature extraction proved more efficient than fine-tuning in terms of training time and resources.

Conclusions:

• Automated GLUT-1 scoring using deep learning offers a highly accurate and objective alternative to manual assessment.
• Deep feature extraction combined with SVM provides a robust method for tumour hypoxia biomarker analysis.
• The developed automated method has the potential to improve the reliability of predicting tumour hypoxia markers in clinical settings.