Multi-cancer analysis of histopathologic MSI screening based on digital histology image

Area of Science:

• Computational pathology
• Genomics
• Artificial intelligence in oncology

Background:

• Microsatellite instability (MSI) is a key biomarker for cancer treatment selection.
• Detecting MSI using digital pathology and deep learning offers a novel approach.

Purpose of the Study:

• To evaluate deep learning models for classifying high microsatellite instability (MSI-H) and microsatellite stable (MSS) using whole-slide images (WSI).
• To determine the most effective pre-trained deep learning model for MSI detection across different cancer types.

Main Methods:

• Whole-slide images (WSI) from public datasets were used to train and evaluate MSI classification models.
• Models were trained on single tissue types (colorectal, stomach, uterine/endometrial cancers) and evaluated on corresponding and cross-tissue datasets.
• A multi-tissue trained model was developed to assess generalizability across cancer types.

Main Results:

• The models achieved high AUC values: 0.93 for colorectal cancer (CRC), 0.84 for stomach adenocarcinoma (STAD), and 0.79 for uterine corpus and endometrial adenocarcinoma (UCEC).
• Models trained on specific tumor tissues showed superior accuracy compared to those trained on different tissues.
• The multi-tissue trained model demonstrated potential for generalizable feature detection across diverse cancer types, though outcomes varied by cancer type.

Conclusions:

• Deep learning models can effectively detect MSI status from WSI, aiding in treatment selection.
• Tissue-specific models offer high accuracy, while multi-tissue models show promise for broader applicability.
• Further research into multi-tissue models could enhance generalizability for MSI detection in various cancers.