Automatic Recognition and Prognostic Prediction of Colorectal Liver Metastases Using a Multi-Scale Deep Learning...

Area of Science:

• Artificial Intelligence in Pathology
• Deep Learning for Medical Imaging
• Computational Pathology

Background:

• Colorectal cancer liver metastasis (CRLM) diagnosis and prognosis face challenges due to limitations in conventional methods.
• Deep learning offers potential for automated feature extraction, improving diagnostic accuracy and prognostic reliability.

Purpose of the Study:

• To develop and validate CLM-Net, a multi-model ensemble deep learning framework for automatic CRLM recognition and prognostic prediction.
• Enhance diagnostic accuracy and clinical applicability of CRLM assessment.

Main Methods:

• Developed CLM-Net using VGG16, DeepLab-v3, and U-Net architectures with multi-scale atrous convolutions and attention mechanisms.
• Trained and validated on 197 CRLM cases from public datasets, employing 5-fold cross-validation and independent testing.
• Evaluated classification, segmentation, and survival prediction using logistic regression, random forest, Kaplan-Meier curves, and log-rank tests.

Main Results:

• CLM-Net achieved 94% accuracy, 92% recall, 93% F1-score, and an AUC of 0.96 for image recognition on an independent test set.
• Survival prediction achieved an AUC of 0.864, outperforming single models in risk stratification.
• Pathologist evaluation showed 90% concordance, indicating strong interpretability and clinical utility.

Conclusions:

• The CLM-Net framework significantly enhances CRLM recognition and prognostic prediction through integrated deep learning approaches.
• Demonstrated excellent generalization and potential for clinical translation in CRLM management.
• Future work includes multi-center validation and multimodal feature integration for precision medicine applications.