Transfer Learning Approach and Nucleus Segmentation with MedCLNet Colon Cancer Database

Area of Science:

• Computational pathology
• Medical image analysis
• Artificial intelligence in oncology

Background:

• Machine learning, particularly deep learning, offers potential for efficient and accurate cancer diagnosis by reducing expert workload.
• Nucleus segmentation in histopathology images is crucial for disease detection and parameter assessment.
• Existing segmentation algorithms like graph theory, PSO, watershed, and random walker have been applied to nucleus detection.

Purpose of the Study:

• To propose and evaluate a transfer learning technique for nucleus segmentation and classification in colorectal cancer histopathology images.
• To introduce the MedCLNet visual dataset for transfer learning studies in deep learning.
• To improve the accuracy and efficiency of colon cancer detection through automated analysis.

Main Methods:

• Nucleus segmentation was performed on the colorectal histology MNIST dataset using graph theory, PSO, watershed, and random walker algorithms.
• A 10-class MedCLNet dataset was created, comprising NCT-CRC-HE-100K, LC25000, and GlaS datasets, for transfer learning.
• Deep neural networks (DenseNet201, DenseNet169, InceptionResNetV2, InceptionV3, ResNet152V2, ResNet101V2, Xception) were pre-trained using the MedCLNet database and applied to the colorectal histology MNIST dataset for classification.

Main Results:

• The DenseNet169 model achieved 94.29% accuracy without transfer learning and 95.00% accuracy after applying the proposed transfer learning method on the colorectal histology MNIST dataset.
• Analysis before and after transfer learning, including DenseNet169+SVM and DenseNet169+GRU, demonstrated the effectiveness of the proposed approach.
• The study confirmed satisfactory outcomes compared to existing empirical studies, highlighting the potential of the transfer learning method.

Conclusions:

• The proposed transfer learning technique using the MedCLNet database significantly improves nucleus segmentation and classification accuracy for colon cancer detection.
• The developed MedCLNet dataset provides a valuable resource for deep learning-based transfer learning studies in histopathology.
• This automated approach is expected to serve as a valuable secondary evaluation tool for physicians in diagnosing colon cancer.