Hepatocellular carcinoma (HCC) is one of the most common and lethal malignancies worldwide, ranking as the third leading cause of cancer-related deaths. Despite advancements in diagnostic and therapeutic approaches, the prognosis for patients with advanced HCC remains poor, with five-year survival rates as low as 18%. A major contributing factor to this is the substantial heterogeneity observed at the genomic, molecular and histological levels. One of the most clinically significant vascular patterns identified in HCC is the presence of vessels encapsulating tumour clusters (VETC), which has been associated with increased metastatic potential and poorer survival outcomes. Research has also indicated that the VETC pattern may serve as a predictor of response to sorafenib treatment. However, traditional methods for identifying VETC require histopathological examination following surgery, limiting its use for preoperative decision-making.
Deep learning (DL) radiopathomics, an emerging field that combines radiomics and pathomics, offers a novel approach to predicting tumour characteristics non-invasively. Radiomics extracts quantitative features from medical images, while pathomics analyses digitised histopathological slides to provide deeper insights into tissue composition. A recent study aimed to develop and validate DL models utilising contrast-enhanced MRI and pathomics to predict VETC and assess the prognosis of patients with HCC. By leveraging deep learning techniques, the study sought to enhance risk stratification and improve clinical decision-making for patients undergoing treatment.
Development of Deep Learning Models for VETC Prediction
A retrospective, multicentre study involving 578 patients with HCC was conducted to develop DL radiopathomics models capable of predicting the presence of VETC. The study population was divided into training (n = 317), internal validation (n = 137) and external validation (n = 124) datasets to ensure robust model performance. The researchers applied four deep learning architectures—ResNet50, DenseNet121, Vision Transformer and Swin Transformer—using transfer learning techniques to classify tumour images. Tumour segmentation was performed manually on contrast-enhanced MRI scans, focusing on arterial, portal venous, hepatobiliary and combined phases. Pathomic features were extracted from haematoxylin-eosin-stained histological slides, with both handcrafted and deep learning-derived features assessed.
Must Read: Advancing HCC Diagnosis with Combined Imaging
Among the tested models, the Swin Transformer achieved the highest predictive performance, with AUC values ranging from 0.77 to 0.79 in radiomics-based predictions and 0.79 in pathomics-based predictions in the external test dataset. Patients with VETC-positive HCC exhibited significantly higher deep radiomics and pathomics scores than those without the VETC pattern. Gradient-weighted class activation mapping (Grad-CAM) heatmaps further confirmed that the deep learning models focused on key tumour regions, particularly the tumour margin, which is known to play a critical role in cancer invasion and metastasis. These findings support the use of DL radiopathomics models as a promising tool for non-invasive VETC prediction.
Prognostic Value of Radiopathomics Models
In addition to predicting VETC status, the models were evaluated for their prognostic utility in stratifying patients based on the risk of early recurrence and progression-free survival (PFS). A radiopathomics nomogram was constructed by incorporating deep radiomics scores, handcrafted and deep pathomics features and relevant clinical variables. The prognostic performance of the nomogram was assessed using concordance indices and time-dependent receiver operating characteristic (ROC) curves.
The model demonstrated concordance indices of 0.69, 0.60 and 0.67 for early recurrence in the training, internal and external test sets, respectively. Time-dependent AUC values for three-year PFS prediction were 0.83, 0.81 and 0.78 across the respective datasets. Patients categorised as high-risk by the model exhibited significantly shorter recurrence-free intervals and reduced PFS compared with those in the low-risk group. The integration of DL-derived features with clinical data enabled the model to more accurately stratify patients according to their likelihood of disease progression. These findings highlight the potential of radiopathomics in refining risk assessment and guiding treatment strategies.
Clinical Implications and Future Directions
The development of transformer-based deep learning models represents a significant advancement in the field of medical imaging and precision oncology. Compared with conventional convolutional neural networks (CNNs), transformer architectures offer an improved ability to capture global image context, leading to enhanced predictive accuracy. In this study, Swin Transformer models outperformed traditional CNNs, achieving higher AUC values for both radiomics and pathomics-based predictions. The results underscore the feasibility of integrating multiphase MRI with histopathological imaging to capture tumour heterogeneity more effectively.
While the study demonstrated strong predictive and prognostic performance, several challenges remain. The retrospective nature of the study introduces potential selection bias, and the relatively short follow-up period limits the assessment of overall survival. Furthermore, pathomics features were derived from representative histological images rather than whole-slide imaging, which may not fully capture the entire tumour microenvironment. Future research should focus on expanding datasets, incorporating multi-modal imaging techniques and conducting prospective clinical trials to validate the generalisability of these models. Additionally, improving model interpretability and integrating deep learning outputs into clinical workflows will be critical for real-world implementation.
DL radiopathomics models provide a promising non-invasive approach for predicting VETC patterns and assessing prognosis in HCC. By integrating MRI-derived radiomics with histology-based pathomics, the models achieve high predictive accuracy and enable improved risk stratification. The findings support the adoption of artificial intelligence-driven tools in oncological decision-making, with further validation required to facilitate their clinical translation. As deep learning technology continues to evolve, its integration into routine practice may contribute to more personalised and effective treatment strategies for patients with HCC.
Source: Radiology: Imaging Cancer
Image Credit: iStock
