Work

Histopathological images-based deep-learning for risk stratification of uveal melanoma

Year: 2025

Type: article

Abstract: To the Editor: Uveal melanoma (UVM) is a rare yet highly aggressive cancer originating in the uvea, the vascular layer of the eye, encompassing the iris, ciliary body, and choroid.[1,2] Although UVM has a lower incidence compared to other melanomas, it carries a high mortality rate due to its tendency to metastasize.[3] The liver is the most common site of metastasis for UVM, and once the cancer has spread to the liver, the prognosis worsens significantly due to limited treatment options.[4] Early detection and a thorough understanding of UVM and its metastatic patterns are crucial for improving patient outcomes.[5] To address these challenges, we developed a comprehensive deep-learning network for diagnosing histological types and predicting liver metastasis in patients with UVM. Moreover, we created a deep-learning Cox model for risk classification, conducted bioinformatic analyses to explore the biological characteristics of subtypes, and constructed a prognostic gene signature validated across multiple cohorts. Furthermore, a nomogram model was developed by integrating clinical and molecular variables [Figure 1].Figure 1: The overall study diagram for the current study. (A) The detail flow work for construction of ensemble deep-learning network. (B) The flow chart of bioinformatic analysis and experimental validation. AUC: Area under the curve; CV: Cross-validation; DGE: Differential gene expression; DL: Deep-learning; LASSO: Least absolute shrinkage and selection operator; KM: Kaplan–Meier; RF: Random forest; WSIs: Whole slide images.Herein, we analyzed 80 whole-slide images (WSIs) from The Cancer Genome Atlas Uveal Melanoma (TCGA-UVM) and a validation cohort of 68 UVM patients from West China Hospital (HX-UVM) in Chengdu, China. WSIs of patients with UVM receiving radical enucleation were obtained with their agreement and written informed consent was acquired. The West China Hospital Ethics Committee approved the study (No. 20242000), which adhered to the principles of the Declaration of Helsinki. The WSIs served as the foundation for studying histological types and liver metastasis in patients with UVM. In addition, RNA sequencing (RNA-seq) data and clinical characteristics from 250 samples across five open-access UVM cohorts (E-MTAB-4097, GSE22138, GSE27831, GSE44295, and GSE84976) were obtained. The TCGA-UVM cohort was randomly divided into a training dataset (70%) and a test dataset (30%). The training set was used to develop the model and fine-tune hyperparameters, whereas the test set and HX-UVM dataset were used to evaluate the model’s prediction performance. Data normalization and augmentation techniques, such as horizontal flip, vertical flip, and random rotation, were used to enhance training data diversity and quality. Using a weakly supervised method with histological types for supervision, a ResNet50 model was trained for 20 epochs with an stochastic gradient descent (SGD) optimizer, a learning rate of 10-2, and L2 regularization. A classifier within the ResNet50 architecture assigned labels to tiles, generating a heatmap of probability predictions across the WSI. These probability tiles created a heatmap representing the probability distribution across the entire WSI, and deep-learning (DL) features were estimated by analyzing the histogram of tile probability scores. Differentially expressed DL features were first identified using the Wilcoxon or Kruskal–Wallis test. A Lasso regression or Random-Forest model was applied to refine and remove redundant DL features. Finally, 10 traditional machine learning models were trained using a selected subset of features and evaluated through 10-fold cross-validation to construct a robust predictive deep-learning network for diagnosing histological types and predicting liver metastatic cases based on DL features. Detailed methods and statistical analysis are shown in Supplementary Materials, https://links.lww.com/CM9/C538. The performance of the histological classifier for diagnosing histological types (Epithelioid, Mixed, and Spindle) in patients with UVM was rigorously validated using the TCGA-UVM dataset and the HX-UVM cohort. At the tile level, the classifier’s performance was assessed with various metrics and visualizations. Receiver operating characteristic curves (ROC) and Precision-Recall curves (PRC) showed an area under the curve (AUC) of 1.000, 0.984, and 0.986 for the diagnosis of epithelioid, mixed, and spindle types, respectively. The evaluation extended to the test dataset and HX-UVM cohort, where the classifier demonstrated exceptional performance with a macro-AUC of 0.989 for ROC and a macro-AUC of 0.984 for PRC in the test dataset. In the HX-UVM cohort, the model achieved a macro-AUC of 0.992 for ROC and a macro-AUC of 0.983 for PRC [Supplementary Figure 1, https://links.lww.com/CM9/C538]. A comprehensive heatmap was created at the WSI level by integrating multiple probable tiles. Gradient-weighted Class Activation Mapping (Grad-CAM) was used to visualize the heatmaps of the final convolutional layers for the deep learning model. Heatmaps overlaid onto the original images highlighted the areas where the model focused [Supplementary Figure 2, https://links.lww.com/CM9/C538]. A histogram of tile likelihoods was then used to assess the entire slide, extracting 101 DL features. Through the Kruskal–Wallis test and false discovery rate (FDR) adjustment, 61 DL features were selected for further analysis. The top 10 features, identified using Gini importance, were used in 10 machine learning algorithms and validated with 10-fold cross-validation. XGBoost, AdaBoost, GradientBoosting, and NaiveBayes classifiers achieved the highest accuracy. The XGBoost model accurately distinguished histological types at the WSI level across the training, test, and HX-UVM datasets [Supplementary Figure 3, https://links.lww.com/CM9/C538]. The primary cause of death in UVM is distant liver metastasis, which is closely linked to the histopathological subtype. A predictive model for liver metastasis was constructed using DL features. The Wilcoxon test and FDR adjustment selected 57 DL features with differential expression in UVM patients with liver metastasis. Lasso regression further reduced the dimensionality, identifying six significant candidate variables. These features were validated using 10 machine learning algorithms, with the Extra-Trees classifier being selected for liver metastasis prediction. The ROC and PRC were 0.962 and 0.939 in the training dataset, 0.826 and 0.102 in the test dataset, and 0.860 and 0.288 in the HX-UVM dataset [Supplementary Figure 4, https://links.lww.com/CM9/C538]. A DL Cox risk model was developed using six DL features associated with liver metastasis. The risk heatmap in the TCGA-UVM cohort illustrated the distribution of risk scores, survival status, and DL feature scores. Patients were categorized into high-risk and low-risk subtypes based on the optimal cutoff value. Survival analysis revealed poorer survival rates among high-risk patients. The ROC curves showed AUC values of 0.589, 0.679, and 0.686 at 1-year, 3-year, and 5-year intervals, respectively. In the HX-UVM cohort, patients were similarly stratified, with survival curves indicating a worse prognosis for high-risk individuals. Gene Set Variation Analysis (GSVA) assessed differential expression patterns across 50 cancer hallmark pathways. Twelve pathways showed differential expression between the two subtypes. The chi-squared test revealed significant differences in clinical parameters such as vital status, metastasis, chromosome 3 status, and stage between the subtypes. Immune characteristics were explored, generating a comprehensive heatmap of 22 types of immune cells, immune-related signatures, and immune-checkpoint genes. Wilcoxon tests indicated that only cluster of differentiation (CD) 4 memory-activated T cells and NK resting cells were differentially expressed. The “maftools” package highlighted subtype-specific mutation patterns, including notable mutated genes such as GNAQ, GNA11, and SF3B1. Significant differences were also observed in signature 2 (apolipoprotein B editing catalytic polypeptide-like mutation [APOBEC]-mediated activities), the APOBEC signature, and microsatellite instability (MSI) [Supplementary Figure 5, https://links.lww.com/CM9/C538]. Next, the Connectivity Map (CMap) was used to predict potential drug treatments for UVM. Gene expression profiles of the 150 most upregulated and 150 most downregulated genes were analyzed using the CMap database. This analysis identified five potential drugs with the lowest scores: mercaptopurine, arachidonyltrifluoromethane, STOCK1N.35696, exisulind, and tacrolimus. A gene risk signature was then developed to guide UVM prognosis and treatment strategies. The TCGA-UVM cohort was used as the training set, with additional validation from other datasets. Patients were categorized into high-risk and low-risk subgroups using a DL signature. Differential expression analysis identified 29 differentially expressed genes (DEGs), with 10 genes significantly associated with survival. Through Lasso-penalized selection, nine genes with robust predictive power were identified. A six-gene signature (CCND3, ENTPD6, IL17RC, KCNK2, RGS2, and SEMA3C) achieved the highest AUC value of 0.832, demonstrating its prognostic potential. Validation in independent cohorts yielded consistent results. A meta-analysis confirmed the gene signature as a significant risk factor influencing UVM survival, with a hazard ratio of 3.58 [Supplementary Figure 6, https://links.lww.com/CM9/C538]. The relative expression of the six-gene signature was further examined in Mel270 (primary) and Omm2.3 (metastatic) cell lines, revealing significant differences (primer sequences are listed in Supplementary Table 1, https://links.lww.com/CM9/C538). In addition, 6-mercaptopurine and arachidonyltrifluoromethane (AACOCF3) were found to inhibit the proliferation of UVM cell lines. The wound-healing assay demonstrated significant inhibition of Mel270 and Omm2.3 cell invasion. Transwell invasion experiments confirmed that AACOCF3 treatment significantly reduced cell invasion [Supplementary Figure 7, https://links.lww.com/CM9/C538]. An accurate nomogram model was developed for predicting UVM prognosis. Univariate Cox analysis revealed significant associations between the DL signature, gene signature, age, histology type, stage, chromosome 3 status, and metastasis with overall survival (OS). Meanwhile, multivariate Cox analysis identified age, stage, and metastasis as significant factors. A comprehensive nomogram model incorporating these variables estimated the five-year OS probability. The study population was divided into three risk groups, with survival curves demonstrating excellent discriminative ability. Time-dependent ROC curves showed higher AUCs (>0.890) for one-, three-, and five-year predictions. Calibration curves exhibited good consistency between predicted and observed outcomes, and the decision curve analysis indicated higher net benefits compared to stage and metastasis [Supplementary Figure 8, https://links.lww.com/CM9/C538]. Through these analyses, a comprehensive deep-learning network and gene signature model were successfully constructed for diagnosing histological types and predicting liver metastasis in UVM. Experimental drug efficacy studies further validated these models. In summary, applying deep-learning models in analyzing hematoxylin and eosin (H&E)-stained histopathological images shows significant potential for enhancing the diagnosis, prognosis, and treatment of UVM patients. These models empower clinicians to make more informed decisions, develop personalized treatment strategies, and ultimately improve patient outcomes. Ongoing research and advancements in deep-learning-based histopathological analysis are expected to advance precision medicine and patient care. Funding This work was supported by the Science and Technology Department of Sichuan Province (No. 2024NSFSC1720) and Chengdu Municipal Science and Technology Bureau Key R&D Support Program (No. 2023-YF09-00041-SN). Conflicts of interest None.

Source: Chinese Medical Journal

DOI: 10.1097/cm9.0000000000003739