Pancreatic cancer, specifically pancreatic ductal adenocarcinoma (PDAC), is a highly lethal malignancy with a dismal prognosis, characterized by a 5-year survival rate of only 6%. The lack of reliable prognostic markers and effective targeted therapies further complicates its management. Genetic mutations, particularly in KRAS, p53, CDKN2A, and SMAD4/DPC4, play a crucial role in PDAC development and progression. KRAS mutations are prevalent in over 90% of cases, while p53 mutations occur in 50-70%, highlighting their significance. p53, a tumor suppressor gene, when mutated, contributes to uncontrolled cell growth and poor prognosis. PD-L1, a biomarker for immune checkpoint inhibitor (ICI) therapies, has gained attention as a potential therapeutic target. High PD-L1 expression is associated with worse prognosis in various cancers and is being explored in the context of ICI effectiveness. The challenge lies in the invasive nature of obtaining tissue samples for genomic analysis. Radiogenomics, a field combining radiomics (quantitative image analysis) and genomics, offers a non-invasive alternative. By extracting quantitative imaging features from routinely acquired CT or MRI scans, radiogenomics aims to predict molecular profiles, potentially guiding personalized treatment strategies. This study aimed to explore the feasibility of using radiogenomics to predict p53 status and PD-L1 expression, ultimately improving prognostication and treatment decisions in PDAC patients. The non-invasive and cost-effective nature of this approach is particularly attractive given the current limitations of traditional biopsy-based genomic profiling.

Extensive research has established the critical role of genetic alterations in pancreatic cancer development and progression. KRAS mutations are almost ubiquitous in PDAC, while mutations in tumor suppressor genes like p53, CDKN2A, and SMAD4 occur frequently. While KRAS mutations are well-established, the prognostic significance of p53 mutations is more complex and has been subject to debate. Some studies have correlated p53 mutations with poor prognosis, while others have not demonstrated a clear association. PD-L1 expression has emerged as a significant biomarker, especially in the context of immunotherapy. Its expression in tumor cells is often associated with an immunosuppressive microenvironment and has been linked to poor prognosis in various cancers, including pancreatic cancer. The use of immune checkpoint inhibitors targeting PD-L1 is showing promise in treating several cancer types, and there is considerable interest in its application to pancreatic cancer. However, the variability and limitations of PD-L1 expression assays necessitate further investigation into reliable predictive biomarkers. The field of radiogenomics has shown promising results in predicting molecular profiles from medical images in other cancers. However, its application to predicting genetic markers such as p53 and PD-L1 in pancreatic cancer is less explored, presenting a significant research opportunity.

This retrospective study included 107 patients diagnosed with PDAC between January 2013 and December 2017 who had not received preoperative chemotherapy. Immunohistochemistry (IHC) was performed on tissue samples to determine p53 and PD-L1 expression. p53 status was defined as positive if the percentage of stained tumor cell nuclei exceeded 20% or was completely absent. PD-L1 status was considered positive if the percentage of tumor cells with membrane staining was ≥1%. Two-phase contrast-enhanced CT scans were acquired for all patients. Two independent experts (radiologist and surgeon) delineated the volume of interest (VOI) encompassing the tumor (VOIpc) and a 4mm expanded peritumoral region (VOI+4mm) on both early and late-phase images. Imaging features (IFs) were extracted using PyRadiomics v2.2.0, resulting in a total of 2074 features (1037 features from each phase for each VOI). Feature selection was performed using Mann-Whitney U test followed by recursive feature elimination with random forest. The selected features were then used to train XGBoost machine learning models to predict p53 and PD-L1 status. Model performance was evaluated using nested cross-validation (inner 5-repeat 5-fold and outer 10-repeat 5-fold) and ROC analysis. The relationship between clinicopathological factors and prognosis was assessed using the Kaplan-Meier method and log-rank tests. Multivariate Cox regression analysis was conducted to identify independent prognostic factors.

The study found that p53 and PD-L1 expression were significant independent prognostic factors (P=0.008 and P=0.013, respectively). 75 (70%) patients were p53-positive, and 36 (33.6%) patients were PD-L1-positive by IHC. The AUC of the predictive models were 0.795 for p53 and 0.683 for PD-L1, using features from the VOI+4mm. Radiogenomics-predicted p53 mutations showed a significant association with poor prognosis (P=0.015). Conversely, the predicted PD-L1 expression did not show a significant correlation with prognosis (P=0.096). Multivariate analysis revealed that p53, PD-L1, CA19-9, surgical procedure, and lymph node metastasis were independent predictors of survival. Other significant factors associated with worse prognosis included high CA19-9 levels, longer operation time, and venous and nerve invasion.

This study demonstrates the potential of radiogenomics in predicting p53 mutations in PDAC, offering a non-invasive approach to assess tumor genetics. The predictive model showed good accuracy for p53 status, and the predicted p53 status was significantly associated with prognosis, supporting the clinical relevance of this approach. The model's performance for PD-L1 was modest, potentially due to the inherent complexity of PD-L1 expression regulation and the limitations of using only CT imaging data. Further research is warranted to explore the integration of multi-modal imaging data (e.g., MRI and PET) to enhance prediction accuracy. The identification of p53 status through radiogenomics holds potential in guiding treatment decisions, particularly in the development of targeted therapies focusing on restoring p53 function. Future research should explore the translation of this radiogenomic approach into clinical practice, including validation in larger, independent cohorts and integration with other genomic and clinical factors to refine prognostication and treatment strategies.

This study highlights the promising utility of radiogenomics for predicting p53 status in PDAC, providing a non-invasive method for assessing a key genetic marker associated with prognosis. While the prediction accuracy for PD-L1 was less robust, the findings suggest that radiogenomics can contribute to personalized medicine for PDAC by enabling non-invasive prediction of critical genetic information. Future studies should focus on validating these findings in larger, independent cohorts and exploring the integration of multi-modal imaging data to improve predictive power. Further investigation into the development of targeted therapies based on radiogenomically predicted p53 status is also warranted.

This study is a retrospective analysis with a relatively small sample size. The study relied solely on CT scans, and the inclusion of additional imaging modalities like MRI and PET might improve prediction accuracy. The selection of features for the machine learning model could influence the results, and external validation in independent datasets is crucial. The pathologists agreement was at 80% and there is a need for additional studies with larger dataset for increased reliability.