Automatic Recognition and Prognostic Prediction of Colorectal Liver Metastases Using a Multi-Scale Deep Learning Framework: Model Development and Validation Study.

Background: Colorectal cancer liver metastasis (CRLM) presents considerable challenges in both diagnosis and prognosis, as conventional approaches often are limited by subjectivity, variability, and limited efficiency. Recent advances in deep learning have shown great potential for automated extraction of pathological features, offering improved diagnostic accuracy and more reliable prognostic predictions. Objective: This study aimed to develop and validate a multi-model ensemble deep learning framework (colon cancer liver metastasis network [CLM-Net]) for the automatic recognition and prognostic prediction of CRLM from pathological images, thereby enhancing diagnostic accuracy and clinical applicability. Methods: A total of 197 pathologically annotated CRLM cases were collected and integrated from publicly available datasets (Kaggle and The Cancer Imaging Archive) to construct high-quality training and independent test sets. CLM-Net was built upon base convolutional neural network architectures including VGG16, DeepLab-v3, and U-Net, incorporating multi-scale atrous convolutions, a squeeze-and-excitation attention mechanism, a conditional random field refinement module, and transfer learning strategies. The framework was comprehensively evaluated using 5-fold cross-validation and independent testing for classification, segmentation, and prognostic prediction tasks. For survival prediction, 1024-dimensional image feature vectors extracted by CLM-Net were analyzed using logistic regression and random forest classifiers, with Kaplan-Meier curves and log-rank tests used for survival analysis. Results: CLM-Net demonstrated superior performance in pathological image recognition, achieving 94% accuracy, 92% recall, 93% F1-score, and an area under the receiver operating characteristic curve of 0.96 on the independent test set, outperforming single models. For survival prediction, CLM-Net with multi-scale attention achieved an area under the receiver operating characteristic curve of 0.864. Kaplan-Meier analysis revealed its significantly stronger risk stratification ability compared with VGG16, U-Net, and DeepLab-v3. Precision-recall curves and heatmaps further confirmed the model's high robustness and generalizability in unseen samples. Clinical evaluation by pathologists indicated strong interpretability and diagnostic utility, with a concordance rate of 90%. Conclusions: The proposed CLM-Net framework, through the integration of diverse deep learning architectures and mechanisms, markedly improved the recognition and prognostic prediction of CRLM, demonstrating excellent generalization and clinical translation potential. Future studies will focus on multi-center validation and integration of multimodal features to further optimize its role in precision medicine.