Detecting microsatellite instability in colorectal cancer using Transformer-based colonoscopy image classification...

Area of Science:

• Oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Colorectal cancer (CRC) poses a significant global health challenge.
• Microsatellite instability-high (MSI-H) is a key feature in certain CRCs, associated with a better prognosis.
• Current MSI status determination often relies on histopathology, necessitating alternative methods.

Purpose of the Study:

• To develop a non-invasive prediction model for MSI status in colorectal cancer using colonoscopy images.
• To evaluate the efficacy of vision Transformer (ViT) models compared to traditional convolutional neural networks (CNNs) for MSI prediction.
• To explore the potential of content-based image retrieval (CBIR) for clinical decision support in CRC diagnosis.

Main Methods:

• Utilized a dataset of 427 MSI-H and 1590 MSS colonoscopy images.
• Employed vision Transformer (ViT) models with pre-trained features for MSI prediction.
• Compared ViT performance against DenseNet201 using support vector machine (SVM) and content-based image retrieval (CBIR).

Main Results:

• The ViT model achieved 84% accuracy and an AUC of 0.86, outperforming DenseNet201 (80% accuracy, 0.80 AUC).
• ViT features demonstrated superior performance in CBIR tasks, with a mean average precision of 0.81 versus 0.79 for DenseNet201.
• ViT mitigated common CNN limitations like limited receptive fields and gradient disappearance.

Conclusions:

• Vision Transformer models show significant promise for accurate, non-invasive MSI status prediction in colorectal cancer from colonoscopy images.
• The ViT-based approach offers a potentially more cost-effective and accessible alternative to histopathology for MSI determination.
• CBIR using ViT features can provide compelling deep learning-driven insights for clinical application in CRC diagnostics.