Deep transfer learning methods for colon cancer classification in confocal laser microscopy images

Area of Science:

• Oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Histological evaluation is the gold standard for detecting peritoneal colorectal cancer metastases.
• Confocal laser microscopy (CLM) offers real-time in vivo imaging for tissue differentiation.
• Automated image classification could enhance surgical workflows by providing immediate feedback.

Purpose of the Study:

• To assess the feasibility of classifying colorectal cancer tissue using CLM images.
• To investigate the effectiveness of convolutional neural networks (CNNs) and transfer learning for this classification task.
• To differentiate between benign and malignant tissues, as well as various tissue types.

Main Methods:

• Analysis of CLM images from the colon and peritoneum.
• Application of classical and state-of-the-art CNNs for direct image learning.
• Investigation of transfer learning strategies, including partial freezing and full fine-tuning, due to a small dataset.

Main Results:

• High performance in classifying peritoneal metastases (AUC of 97.1) and colon primary tumors (AUC of 73.1).
• Transfer learning significantly outperformed training models from scratch.
• The optimal transfer learning strategy was found to be model- and task-dependent.

Conclusions:

• CNNs combined with transfer learning can effectively identify cancerous tissue via CLM.
• Task-specific model and transfer learning strategy selection is crucial for optimal performance.
• The high accuracy in the peritoneum suggests potential for intraoperative decision support, even with limited data.