This study was conducted using publicly available datasets, including the Liver Tumor Segmentation dataset from Kaggle and The Cancer Imaging Archive (TCIA). Because the data are publicly accessible and fully deidentified, institutional review board approval was not required for this retrospective study in accordance with institutional policies of Yongkang First People’s Hospital Affiliated to Hangzhou Medical College (policy document available upon request). The study was conducted in accordance with the ethical principles outlined in the Declaration of Helsinki.

All datasets used in this study contain anonymized patient information, and no identifiable personal data were accessed. Therefore, informed consent from individual participants was not required.

The study adhered to principles of data privacy and confidentiality, as only deidentified imaging and clinical metadata were used for research purposes.

No financial or other compensation was provided to participants, as this study did not involve direct patient recruitment or intervention.

The pathological image data used in this study were primarily derived from 2 public databases: the Liver Tumor Segmentation Challenge dataset on Kaggle and TCIA. The Kaggle dataset contained 129 CRLM-related histopathological slides with corresponding clinical information, including age, sex, and tumor stage. To further enhance the sample size and diversity, an additional 68 CRLM cases were obtained from TCIA. In total, image data from 197 patients were included. After quality control and image augmentation, a final set of 1000 high-quality images was generated for model training and evaluation.

During preprocessing, low-resolution or unevenly stained images were excluded. All images were resized to 256×256 pixels to meet the input requirements of the model. Annotations were performed by experienced pathologists to ensure the accuracy of metastatic lesions and surrounding regions. Each image was linked to a unique patient identifier to guarantee patient-level independence across training, validation, testing, and prognostic modeling tasks.

Pathological image annotation was conducted by board-certified pathologists with over 10 years of professional experience. Using Aperio ImageScope software (version 12.3.3; Leica Biosystems, Germany), the experts meticulously annotated tumor regions and liver metastasis features within each slide. In addition, they recorded relevant clinical parameters, including tumor size and morphology, clarity of tumor boundaries with adjacent normal liver tissue, the spatial distribution of tumor-associated macrophages, and the density and localization of immune cells (eg, T cells and B cells) within the tumor microenvironment. All annotation data were saved in XML format to ensure compatibility and efficient parsing by the subsequent deep learning models.

To ensure annotation accuracy, all pathological images in this study were independently annotated by 2 experienced pathologists. Interrater reliability was assessed using the Cohen κ coefficient. In cases of disagreement, a third senior pathologist reviewed and confirmed the final annotation, ensuring both precision and consistency. To minimize manual annotation errors, the following quality control strategies were implemented:

These measures collectively ensured high-quality annotations, establishing a robust and reliable dataset for subsequent deep learning model training and analysis.

To improve model generalization across diverse pathological image features, various data augmentation strategies were applied, tailored to the specific requirements of each deep learning architecture. For VGG16, images were first resized to 224×224 pixels using bilinear interpolation to match the model’s input specification. Standard normalization was then performed by subtracting the dataset’s mean pixel value and dividing by the SD, thereby transforming the pixel distribution to approximate a standard normal distribution. In the case of DeepLab-v3, numerical data were similarly normalized to a mean of 0 and an SD of 1. For textual data inputs, word embedding techniques were used to convert words into dense vector representations using pretrained models such as Word2Vec or GloVe for initialization. For U-Net, input images were also resized and normalized as described above. In addition, random cropping and horizontal flipping were implemented to increase data variability and reduce overfitting. All preprocessing steps were conducted using standard APIs provided by mainstream deep learning frameworks such as TensorFlow and PyTorch. Specifically, the torchvision.transforms module in PyTorch was used for image transformations.

Several classical CNN models, including DeepLab-v3, VGG16, and U-Net, were used for the classification and segmentation of pathological images. All models were implemented using the PyTorch framework (version 1.7.1; Meta AI, Meta Platforms, Inc). To accelerate convergence and improve generalization, model parameters were initialized with pretrained weights from the ImageNet dataset. Training was conducted using the Adam optimizer, with a learning rate of 1×10^–4, a batch size of 32, and 50 training epochs.

DeepLab-v3 is a state-of-the-art semantic segmentation model that incorporates dilated (atrous) convolutions to capture multi-scale contextual information without reducing spatial resolution. Its core component, atrous spatial pyramid pooling (ASPP), applies parallel atrous convolutions with varying dilation rates to extract rich semantic features at multiple scales, thereby enhancing model adaptability to complex structures in histopathological images. The use of dilated convolutions expands the receptive field without increasing the number of parameters or computational cost, enabling more comprehensive feature extraction. DeepLab-v3 follows an encoder-decoder architecture, where the encoder captures high-level semantic features, and the decoder restores spatial details for precise segmentation. Commonly, pretrained backbone networks such as Xception or MobileNet are adopted to serve as feature extractors, followed by the ASPP module and decoder to complete the segmentation task. The model structure is illustrated in Figure 1.

VGG16 is a classic deep CNN composed of 16 layers, including 13 convolutional layers and 3 fully connected layers. It captures image features by repetitively stacking 3×3 convolutional kernels and 2×2 max-pooling layers. By using multiple small kernels, VGG16 achieves a larger receptive field while maintaining a relatively low number of parameters and a deeper network structure. Its architecture is straightforward and consistent, with fixed strides in all convolutional and pooling layers, which simplifies network design and optimization. In our study, VGG16 was initialized with pretrained weights from the ImageNet dataset. During training, images were preprocessed by mean subtraction and optimized using mini-batch stochastic gradient descent. The algorithmic framework is illustrated in Figure 2.

U-Net is a neural network model specifically designed for medical image segmentation, characterized by its U-shaped architecture consisting of a contracting path (encoder) and an expansive path (decoder). A key feature of U-Net is the use of skip connections that merge encoder feature maps with corresponding decoder outputs, thereby preserving spatial context and enhancing the reconstruction of fine-grained details. U-Net relies heavily on data augmentation techniques during training, such as rotation, scaling, and elastic deformation, to improve generalizability. In our implementation, the model was trained using the Adam optimizer, with a learning rate of 1×10^–4, a batch size of 8, and a cross-entropy loss function. The algorithm structure is shown in Figure 3.

Each of the 3 models demonstrates unique strengths for different image analysis tasks. DeepLab-v3 is suited for fine-grained segmentation and multi-scale feature extraction. VGG16, with its efficient and compact design, is often used as a backbone for various visual recognition tasks. U-Net, in contrast, excels in medical image segmentation, particularly in scenarios with limited datasets. Based on the complementary strengths of these architectures, we developed a customized CNN framework, CLM-Net, specifically tailored for the automatic recognition of pathological features in CRLM.

CLM-Net is a CNN framework specifically developed for the automatic recognition of pathological features in CRLM histopathological images. The architecture adopts a parallel fusion structure that integrates the strengths of 3 widely used CNN models: VGG16, DeepLab-v3, and U-Net. Initially, input images are processed through a VGG16 backbone pretrained on ImageNet to extract low-level texture features. These features are then passed into 2 parallel branches: a DeepLab-v3 path and a U-Net path. The DeepLab-v3 branch applies atrous convolutions with dilation rates of {1, 2, 4, 8}, combined with an ASPP module, to capture rich multi-scale contextual information. Meanwhile, the U-Net branch adopts an encoder–decoder structure with skip connections to preserve spatial resolution and boundary detail. Mid-level feature fusion is performed by channel-wise concatenation of outputs from the VGG16, DeepLab-v3, and U-Net bottleneck layers.

To further enhance feature discrimination, a SE attention module (reduction ratio r=16) is inserted after the fusion stage to adaptively recalibrate channel-wise feature importance, improving the model’s sensitivity to tumor-relevant features while suppressing background noise. The fused features are then passed through a unified decoder, and the final segmentation map is refined using a conditional random field module to enhance pixel-level consistency, particularly at lesion boundaries. In addition to the segmentation output, the decoder also generates a 1024-dimensional feature vector, which is used in downstream survival prediction tasks.

During training, CLM-Net used the same hyperparameters as the base models, including the Adam optimizer, a learning rate of 1×10^–4, a batch size of 32, and 50 training epochs. To address the limited size of the training dataset and improve model generalizability, extensive data augmentation techniques such as rotation, scaling, and elastic deformation were applied. By integrating low-level feature extraction (VGG16), multi-scale semantic context (DeepLab-v3), and spatial detail preservation (U-Net), CLM-Net provides a robust and comprehensive solution for accurate CRLM lesion segmentation and prognosis prediction. The overall framework is illustrated in Figure 4.

To comprehensively assess the performance of various deep learning models for CRLM pathological image recognition, several metrics were used, including accuracy, recall, precision, F₁-score, and the area under the receiver operating characteristic curve (AUC). For prognostic analysis, survival-related assessments such as Kaplan-Meier curves were used. Parameter tuning was performed using a 5-fold cross-validation strategy. All metric computations were implemented in Python using the classification_report and roc_auc_score functions from the Scikit-learn library. Each model’s performance on an independent test set was carefully documented to ensure objective and reliable evaluation.

To evaluate the generalization capability of each model, a k-fold cross-validation approach was adopted, with k set to 5. Specifically, the dataset was randomly partitioned into 5 equal subsets. In each iteration, one subset served as the validation set, while the remaining 4 were used for training. This process was repeated 5 times to ensure each subset was used once for validation. The cross-validation procedure was executed using the KFold class from the Scikit-learn library. This approach enabled a fair and comprehensive assessment of each model’s generalization performance.

Upon completing training, validation, and evaluation, the models were compared based on average performance across all metrics, including accuracy, recall, F₁-score, and AUC. Custom Python scripts were used to facilitate a standardized and transparent comparison process. The model demonstrating the most favorable overall performance was selected as the optimal model. Subsequent to selection, this model underwent further refinement, including hyperparameter tuning and expanded data augmentation strategies, to enhance prediction accuracy and clinical applicability.

All models were trained using the PyTorch 1.7.1 framework, with the Adam optimizer set at a learning rate of 1e−4, a batch size of 32, and for 50 training epochs to ensure adequate learning of data features. A transfer learning strategy was used by initializing model weights with pretrained parameters from the ImageNet dataset, which helped accelerate convergence and improve training stability. To comprehensively assess model performance, several evaluation metrics were used, including accuracy, recall, F₁-score, and the area under the receiver operating characteristic (ROC) curve (AUC). For prognostic assessment, Kaplan-Meier survival curves were generated to evaluate the model’s ability to predict patient outcomes. All models were evaluated using 5-fold cross-validation. In this approach, the dataset was randomly divided into 5 equal subsets; during each iteration, one subset was designated as the validation set, and the remaining 4 served as the training set. This method ensured robust performance evaluation and validated the generalization capability of the models across different data partitions.

To ensure the generalizability of the model and the reliability of the evaluation results, the dataset was obtained from publicly available databases and randomly divided into training, validation, and independent test sets in a 70%‐15%‐15% ratio. The training set was used for model training, the validation set for hyperparameter tuning, and the independent test set for final performance evaluation. The training set included approximately 700 images, the validation set 150 images, and the independent test set 150 images, all of which were annotated by experienced pathologists. A balanced class distribution between liver metastasis (Class 1) and nonmetastasis (Class 0) was maintained across all subsets to ensure consistent class representation during training and fair performance evaluation. Detailed class distribution across subsets is provided in Table S2 in Multimedia Appendix 1. All images were sourced from publicly available CRLM pathology databases, including TCIA and Kaggle. Image quality and annotation accuracy were verified through multiple rounds of manual verification by expert pathologists.

On the independent test set, the CLM-Net model demonstrated outstanding performance. Compared with other baseline CNN models, CLM-Net achieved notably higher scores in accuracy, recall, and F₁-score, indicating stronger generalization and robustness in recognizing pathological features of CRLM.

In this study, the matching rate was defined as the overlap between the model’s predicted positive regions and the ground-truth annotated positive regions, calculated using the intersection over union (IoU). The specific formula is as follows: IoU=Predicted Region∪Ground Truth RegionPredicted Region∩Ground Truth Region.

To evaluate model performance, the IoU value for each pathological image in the independent test set was calculated. The overall matching rate was then evaluated by calculating the average IoU across the entire test set. The experimental results showed that the matching rate for all selected models exceeded 85%, with the CLM-Net model achieving the highest matching rate (close to 90%). This demonstrates the model’s strong regional localization ability. Moreover, the high consistency of the matching rate across different samples reinforces the reliability of CLM-Net in pathological image analysis tasks.

To assess the practical clinical applicability of the model, a standardized survey was designed and conducted. The survey comprised the following components: (1) quantitative assessment of the model’s prediction accuracy, (2) evaluation of usability and user experience through a structured questionnaire, (3) subjective feedback from clinicians regarding the model’s diagnostic assistance capabilities (via open-ended questions), and (4) suggestions for further improvement and optimization of the model (also via open-ended questions). The survey was administered to a cohort of 10 experienced pathologists and clinicians. Questionnaires were distributed electronically and collected during the model validation phase. The responses were analyzed using both quantitative statistical methods and qualitative text analysis, providing critical insights to guide subsequent refinements in model development and optimization.

To assess the prognostic prediction performance of the CLM-Net model in patients with CRLM, both traditional survival analysis methods and machine learning-based classifiers were used. Specifically, Kaplan-Meier survival curves and the Cox proportional hazards model were used to analyze survival outcomes across different patient groups. For prognosis prediction, the 1024-dimensional feature vectors were extracted from the final fusion layer of the CLM-Net encoder. These vectors integrated multi-scale representations from VGG16, DeepLab-v3 (with an ASPP module), and U-Net (with skip connections), further enhanced by an SE attention mechanism. To reduce dimensionality and suppress noise, 2 preprocessing strategies were adopted:

The processed feature vectors were input into 2 classical machine learning models:

For comparative evaluation, the same prediction framework was applied to features extracted from 3 baseline deep learning models—VGG16, DeepLab-v3, and U-Net. Each model generated survival prediction probabilities from the input pathological images. A threshold of 0.5 was used to classify samples into short- and long-term survival groups. Kaplan-Meier curves were plotted for each model using the lifelines Python library, with time on the x-axis and cumulative survival probability on the y-axis. The Log-rank test was used to determine statistically significant differences among survival curves, and the final comparative results are presented in Table S3 in Multimedia Appendix 1.

This study was conducted as a retrospective analysis using publicly available datasets and did not involve any direct clinical trials or the use of patient data beyond those datasets. To assess the practical applicability of the model, we collaborated with clinicians from affiliated hospitals to evaluate its usability and diagnostic accuracy. Clinical experts reviewed the model’s pathological image analysis outputs and compared them with conventional histopathological interpretations to assess its potential value in diagnostic support. Feedback from these professionals was systematically collected, focusing on diagnostic consistency, ease of interpretation, and clinical relevance. Based on this expert feedback, iterative modifications and optimizations were applied to enhance the model’s clinical adaptability and reliability.

Throughout the model development and validation process, data augmentation strategies significantly enhanced the performance of all deep learning models, particularly the ensemble learning model. The application of random rotations, horizontal flips, and cropping effectively increased training data variability, thereby improving model robustness and generalizability to previously unseen pathological images (Figure 5A). This approach is especially critical in the context of CRLM, where metastatic lesions are known to be highly heterogeneous in appearance.

The ensemble learning model (CLM-Net) outperformed individual models across multiple evaluation metrics by leveraging the complementary strengths of each architecture. While certain base models, such as U-Net, excelled at delineating fine-grained tumor boundaries, others (eg, DeepLab-v3) were more proficient in identifying contextual and structural features of metastatic lesions. By integrating these capabilities, CLM-Net achieved a more comprehensive and accurate recognition of CRLM pathological features, thereby enhancing diagnostic reliability (Figure 5B).

In quantitative performance evaluation, CLM-Net demonstrated outstanding results in CRLM pathology image recognition, achieving an accuracy of 94% (95% CI 91.2%‐96.5%), a recall of 92% (95% CI 89.0%‐94.8%), an F₁-score of 93% (95% CI 90.1%‐95.9%), and an AUC of 0.960 (95% CI 0.941‐0.975). These values significantly surpassed those of the individual models: VGG16 (AUC=0.901, 95% CI 0.872‐0.925), DeepLab-v3 (AUC=0.912, 95% CI 0.884‐0.936), and U-Net (AUC=0.887, 95% CI 0.857‐0.913). The AUC improvement of CLM-Net over each baseline was statistically significant (DeLong test, P<.01), confirming its superior capability in both classification and segmentation of CRLM pathological images (Figure 5C). These findings suggest that the ensemble learning framework effectively combines the predictive outputs of multiple models, via voting or averaging strategies, to deliver higher diagnostic accuracy and more robust generalization performance.

In the context of prognostic prediction for patients with CRLM, the integration of multi-scale processing and attention mechanisms demonstrated marked performance advantages. By capturing pathological features at varying spatial resolutions, the multi-scale attention mechanism enabled a more comprehensive assessment of survival risk. Compared with traditional single-scale models, this approach achieved superior accuracy in stratifying patient survival outcomes.

To validate its prognostic effectiveness, survival analysis methods were used. Kaplan-Meier survival curves and the Cox proportional hazards model confirmed the enhanced predictive capabilities of the multi-scale attention framework. Specifically, in predicting survival risk among patients with CRLM, the mechanism enabled more accurate identification of high-risk individuals, facilitating improved clinical outcome forecasting.

CLM-Net, equipped with multi-scale and attention modules, achieved the highest performance in the prognostic task. In ROC curve analysis, CLM-Net reached an AUC of 0.864 (95% CI 0.832‐0.892), significantly outperforming VGG16 (AUC=0.812, 95% CI 0.775‐0.846), DeepLab-v3 (AUC=0.829, 95% CI 0.796‐0.861), and U-Net (AUC=0.794, 95% CI 0.758‐0.827). Differences in AUC were statistically significant according to DeLong test (P<.05) (Figure 6A), supporting the superiority of CLM-Net’s multi-scale design in extracting clinically relevant image features for survival prediction.

Moreover, the integration of attention mechanisms further enhanced the model’s performance. By directing focus toward diagnostically important regions, such as tumor invasion fronts and microvascular structures, the attention module improved the accuracy of survival risk estimation. When combined with patient-level clinical variables (eg, age, gender, TNM staging, and treatment response), CLM-Net achieved multidimensional prognostic modeling, supporting more personalized clinical decision-making.

Comparative Kaplan-Meier survival curve analysis (Figure 6B) further illustrated the model’s superiority. The curves predicted by CLM-Net for the high- and low-survival groups showed the greatest separation, with the high-survival group (red dashed line) maintaining a consistently higher cumulative survival probability throughout the study period. Log-rank tests confirmed the statistical significance of CLM-Net’s predictions compared with those of VGG16 (P=.01), DeepLab-v3 (P=.007), and U-Net (P=.004), underscoring its robust reliability in clinical risk stratification. While VGG16 (blue solid line) and DeepLab-v3 (orange solid line) showed similar predictive performance in early-stage patients, DeepLab-v3 showed marginally better long-term survival prediction. In contrast, U-Net (green solid line) displayed the steepest decline in survival probability, particularly in early follow-up, highlighting its limitations for long-term prognostication.

Further survival analysis (Figure 6C) stratified patients into high-risk (short-term survival) and low-risk (long-term survival) groups using a probability threshold of 0.5 based on model output. CLM-Net’s Kaplan-Meier curves again displayed a clear and consistent separation between the groups, with the log-rank test yielding a highly significant P value (P<.001), confirming the model’s stability and discriminative power. In comparison, survival curves generated by VGG16, DeepLab-v3, and U-Net showed weaker separation, with log-rank P values of .087, .041, and .126, respectively. Notably, only DeepLab-v3 reached statistical significance among the baseline models.

Taken together, these results highlight the clinical utility of the CLM-Net framework in identifying high-risk patients with CRLM. Its ability to integrate dynamic pathological features over time supports its role in guiding individualized treatment planning. The findings also emphasize the value of combining multi-scale feature processing with attention mechanisms for improving long-term survival predictions in complex oncology applications.

To validate the generalizability of the proposed model to previously unseen data, we conducted a systematic evaluation using the independent test set, with a particular focus on the performance of the ensemble learning framework and the multi-scale processing models (Figure 7). In the precision-recall (PR) curve analysis, the ensemble model demonstrated highly robust performance across 2 key tasks—CRLM identification and 28-day survival classification—with the PR curves maintaining a high plateau, indicating strong sensitivity to minority classes (Figure 7A). Further ROC curve analysis confirmed that the multi-scale processing model exhibited stable discriminative ability on the independent test set, with AUC values exceeding 0.95, suggesting that the model maintained high classification accuracy even when applied to novel samples (Figure 7B).

To provide a more intuitive visualization of model performance across different classes and evaluation metrics, we constructed a classification report heatmap based on the independent test set (Figure 7C). This heatmap presents the precision, recall, and F₁-score for Class 0 (nonmetastasis) and Class 1 (metastasis), all exceeding 0.92. The weighted average precision and recall were both 0.93, indicating balanced overall performance with no significant bias. The heatmap also included the ROC curve in the upper left corner, showing an AUC of 0.96, further corroborating the model’s excellent generalizability. Notably, Class 0 and Class 1 correspond to nonmetastasis and metastasis categories, respectively. The stable classification outcomes highlight the model’s strong potential for deployment in real-world clinical practice.

To validate the clinical applicability and prognostic accuracy of the proposed model, comparisons were conducted between model-predicted outcomes and clinical data from publicly available CRLM datasets. Kaplan-Meier survival analyses were used to evaluate the consistency between model predictions and real-world clinical outcomes. In this context, “responders” referred to patients whose clinical progression matched the model’s predictions of favorable outcomes, whereas “non-responders” exhibited discrepancies between predicted and actual clinical results. The analysis revealed a concordance rate exceeding 90% between model-predicted and actual survival outcomes (Figure 8A), demonstrating the model’s strong prognostic reliability in patients with CRLM and underscoring its potential utility in precision oncology.

In addition to quantitative validation, qualitative feedback from physicians provided highly positive feedback on the usability and accuracy of the model’s predictions. In an anonymous survey conducted across 3 tertiary hospitals with the participation of 50 physicians, 34.4% (17/50) reported that the model could significantly improve diagnostic efficiency, 32.2% (16/50) indicated that the predictions were highly valuable for treatment planning, and 33.3% (17/50) considered the model highly reliable in forecasting disease progression and patient survival (Figure 8B). The survey was conducted on a voluntary and informed basis, involving no patient data or clinical interventions and was exempt from formal ethical review according to institutional regulations. The questionnaire content is provided in Table S4 in Multimedia Appendix 1 to enhance transparency and reproducibility.

This study highlights the practical potential of deep learning models in supporting CRLM clinical decision-making. By assisting physicians in evaluating patient prognosis, the model facilitates more informed, efficient, and personalized treatment strategies (Figure 8C). Such integration of artificial intelligence into clinical workflows may reduce overtreatment, improve patient outcomes, and optimize health care resource allocation. Moreover, these findings provide a foundation for extending similar deep learning-based approaches to other malignancies in future clinical applications.

For prognostic modeling, the Cox proportional hazards regression model was used to estimate survival times, while Kaplan-Meier curves were generated to assess survival differences between responders and nonresponders. The inclusion of censored data (ie, right-censored survival times) did not significantly affect the model’s performance, as confirmed by the log-rank test.