Rectal cancer is a significant global health concern, requiring accurate prognostic assessments to guide treatment strategies effectively. While the tumour-node-metastasis (TNM) staging system is commonly used, variations in patient outcomes indicate the need for additional predictive markers. Tumour budding (TB), a histopathological feature associated with poor prognosis, plays a critical role in determining disease progression. TB is defined by the presence of single tumour cells or small clusters of undifferentiated cancer cells at the invasive front of the tumour. It has been linked to increased tumour aggressiveness, lymphatic and vascular invasion and a higher likelihood of metastasis.
Traditionally, TB assessment relies on invasive biopsy procedures, which may not fully capture tumour heterogeneity. Biopsy samples often fail to represent the entire tumour, potentially leading to underestimation of TB grade. The development of noninvasive imaging techniques incorporating artificial intelligence offers a promising alternative for accurate and comprehensive tumour evaluation. A multiparameter deep learning-radiomic model (DLRM) using computed tomography (CT) images and extracellular volume (ECV) parameters has been proposed to predict TB grades in rectal cancer patients, potentially enhancing preoperative decision-making and treatment planning.
Deep Learning and Radiomic Feature Extraction
Radiomics involves the high-throughput extraction of quantitative imaging features, offering insights into tumour characteristics beyond conventional radiology. This method allows for the identification of imaging biomarkers that can reflect tumour heterogeneity, microenvironmental changes and treatment response. Deep learning (DL) techniques enable automated feature extraction from imaging data, identifying complex patterns that may correlate with disease severity. Unlike handcrafted radiomic (HCR) features, DL-derived features are automatically learned from large datasets, reducing potential bias and improving predictive accuracy.
Recommended Read: MRI Radiomics for Lymphovascular Invasion Prediction in Rectal Cancer
In this study, preoperative CT-based ECV parameter images and venous-phase images were analysed using HCR and DL features. The extracted features were then selected using machine learning techniques, including the least absolute shrinkage and selection operator (LASSO) algorithm, to optimise the predictive model. LASSO regression ensures that only the most relevant and non-redundant features are included in the final model, improving its robustness and interpretability. Six machine learning classifiers were developed to construct individual predictive signatures, ultimately leading to the integration of HCR and DL features into the comprehensive DLRM. The combination of these features allows the model to leverage both manually designed and automatically learned patterns, enhancing its overall performance.
Model Performance and Clinical Applicability
The study assessed the predictive performance of the developed model by evaluating its ability to distinguish between low- to intermediate-grade TB (Bd1+2) and high-grade TB (Bd3). The DLRM demonstrated superior performance compared to individual HCR or DL signatures, achieving high area under the receiver operating characteristic curve (AUC) values in both training and test cohorts. The model’s high sensitivity and specificity suggest its potential for reliable preoperative TB grading.
Calibration curve analysis confirmed the model's reliability, while decision curve analysis (DCA) demonstrated its clinical utility by indicating that the DLRM provided greater net benefit compared to alternative approaches. The ability to noninvasively assess TB grade preoperatively has significant clinical implications. By accurately identifying high-risk patients, clinicians can make informed decisions regarding the need for neoadjuvant therapy, surgical planning and post-treatment follow-up. As TB grade has been associated with response to neoadjuvant therapy and overall survival, integrating the DLRM into routine clinical practice could improve personalised treatment approaches and patient outcomes.
Advantages and Limitations
The primary advantage of this approach is its ability to provide a noninvasive, accurate assessment of TB grade, reducing reliance on invasive biopsies. The integration of ECV parameter images enhances model interpretability by reflecting tumour microenvironmental changes, such as alterations in stromal composition and vascularity. Moreover, CT imaging is widely available, making this approach more accessible compared to alternative methods such as MRI-based radiomics, which may have limited applicability due to cost and availability constraints.
However, certain limitations must be considered. The study was retrospective, with a limited sample size, potentially introducing selection bias. Additionally, manual image segmentation may affect feature consistency, necessitating the development of automated segmentation techniques to improve reproducibility. Variability in imaging protocols and scanner settings across different institutions may also impact model performance. Further prospective, multicentre studies are required to validate the model’s generalisability and establish standardised imaging protocols.
Another consideration is the need for further research to evaluate the model’s integration with other prognostic factors, such as molecular and genetic markers. Combining imaging-based predictive models with genomic data may further enhance accuracy and provide a more comprehensive assessment of tumour biology. Additionally, long-term follow-up studies are necessary to determine the model’s prognostic value in predicting disease recurrence and survival outcomes.
The development of a CT-based multiparameter deep learning-radiomic model represents a significant advancement in noninvasive tumour assessment. By integrating deep learning and radiomic features, the DLRM provides a reliable tool for predicting TB grade in rectal cancer patients, aiding in preoperative decision-making. The ability to accurately assess TB grade without requiring invasive procedures offers a transformative approach to patient management, allowing for more precise treatment stratification.
Future research should focus on refining model robustness, expanding patient cohorts and exploring additional imaging modalities to further enhance predictive accuracy and clinical impact. Additionally, efforts should be made to standardise imaging acquisition and processing techniques to ensure reproducibility and facilitate widespread adoption in clinical practice. If validated in larger, prospective studies, this model has the potential to significantly improve the precision of rectal cancer management, ultimately leading to better patient outcomes.
Source: Academic Radiology
Image Credit: iStock
