Machine Learning–Based Enhanced MRI Radiomics for PDCD1 Prognostication and Expression Prediction in Breast Cancer

Abstract

Background:

Programmed cell death 1 (PDCD1) is an immune checkpoint inhibitor that plays an important role in immune evasion in breast cancer (BC). In this study, we aimed to evaluate the correlation between PDCD1 expression, immune cell tumor infiltration, and prognosis. In addition, we also developed a predictive model to determine PDCD1 expression levels in patients with BC based on radiomics features extracted from magnetic resonance imaging (MRI).

Methods:

Clinical data of 1082 patients with BC extracted from The Cancer Genome Atlas (TCGA) and MRI data of 108 patients with BC extracted from The Cancer Imaging Archive (TCIA) were used to determine the correlation between PDCD1 expression levels and the prognosis, clinical stage, survival, and levels of immune cell tumor infiltration in patients with BC. Predictive radiomics features for PDCD1 were extracted by 2 physicians from MRI data. The top 5 predictive features were evaluated and selected to build 2 machine learning models.

Results:

The PDCD1 expression levels were significantly higher in tumor tissues from patients with BC (P < .001). High PDCD1 expression levels were associated with improved overall survival, hazard ratio (HR) = 0.63, 95% confidence interval (CI) 0.425-0.934, P = .021. The PDCD1 expression levels showed a significant positive correlation with immune cell infiltration, including CD8 (P < .001) and Treg (P < .001). Both MRI radiomics models demonstrated good accuracy, strong clinical utility, and a high level of consistency in discriminating between low and high PDCD1 expression levels (P > .05).

Conclusions:

PDCD1 expression showed a good correlation with prognosis and tumor immune cell infiltration. The MRI radiomics model accurately predicted PDCD1 expression levels and could potentially serve as a noninvasive tool to predict early tumor response to immunotherapy.

Keywords

BC MRI radiomics TCGA TCIA machine learning algorithms

Introduction

Breast cancer (BC) is the most common malignant tumor in women and a leading cause of cancer-related mortality worldwide.^1,2 Although surgery, radiotherapy, and chemotherapy have significantly improved treatment outcomes there is a need to develop more personalized and effective treatment approaches.³ Early detection of BC remains crucial because it is essential for improving treatment outcomes. Standard imaging tools such as mammography and magnetic resonance imaging (MRI) continue to play a crucial role in the early detection and management of BC. Traditional prognostic markers such as the Human Epidermal Growth Factor Receptor 2 (HER-2), Kiel-67 (Ki-67), Carcinoembryonic Antigen (CEA), and Cancer Antigen 19-9 (CA19-9) have been indispensable in predicting disease progression and guiding treatment decisions.¹ However, additional biomarkers are required to tailor the treatment according to the needs of the patients and improve outcomes.⁴

Programmed Cell Death 1 (PDCD1) plays a significant role in the immune system’s regulation in several cancers.⁵ PDCD1 is an immune checkpoint receptor expressed on the surface of T cells and other immune cells. When it binds to its ligands, PD-L1 or PD-L2, it transmits an inhibitory signal that reduces the activity of T cells, leading to a decreased immune response which facilitates immune evasion.^6,7 As a result, PDCD1 and its ligands have become important therapeutic targets in BC treatment. Immune checkpoint inhibitors, such as pembrolizumab (anti-PDCD1) and atezolizumab (anti-PD-L1) can block this pathway and restore the T cell’s ability to recognize and kill cancer cells. These therapies have shown promising results in treating certain subtypes of BC, particularly triple-negative BC. The current methods used to detect PDCD1 expression include peripheral blood cytokine assays, mRNA and protein level assessments in fresh tissue specimens, and paraffin tissue specimen analyses using immunohistochemistry or fluorescence techniques. However, these detection methods are invasive, expensive, and cannot be used to capture the expression of PDCD1 within the tumor in real time. Moreover, the accuracy of mRNA and protein analysis may be influenced by operator variability and antibody specificity. Given the constraints of existing detection methods, there is a need to explore noninvasive and efficient approaches to predict PDCD1 gene expression.

Radiomics is a new noninvasive method that is increasingly being used to characterize tumors based on large amounts of quantitative imaging features extracted from radiological images. These features are then used to predict treatment outcomes and provide personalized patient management in oncology and other medical fields.^1,8-13 Previous studies have shown that radiomics features extracted from MRI could be used to facilitate the early characterization of BC lesions and assess lymph node involvement, tumor heterogeneity and the microenvironment.^14-16

Therefore, in this study, we aimed to make use of data extracted from The Cancer Imaging Archive (TCIA) to evaluate the correlation between PDCD1 expression levels, tumor immune cell infiltration, and prognosis in patients with BC. In addition, we also evaluated the ability of machine learning models based on MRI radiomics features in predicting PDCD1 expression levels in patients with BC. The outcomes of this study could provide clinicians with an alternative noninvasive tool to predict PDCD1 expression levels and guide treatment interventions.

Methods

Data collection from The Cancer Genome Atlas and TCIA database

Multi-modal data acquisition and processing were conducted through rigorous inclusion criteria. These criteria include inclusion of only primary, first-diagnosed, treatment-naïve patients with breast cancer (BC); exclusion of patients with missing survival data, follow-up time less than 30 days, male gender, T-stage Tx, or missing clinical variables; and final selection of primary solid tumor samples with RNA-seq data. For imaging data, we excluded post-surgical cases and those with poor-quality images, retaining only patients with high-quality imaging data matched to The Cancer Genome Atlas (TCGA) data for subsequent analysis. Magnetic resonance imaging (MRI) datasets and RNA-sequencing profiles for BC were independently retrieved from TCIA (n = 137) and TCGA (n = 1098) repositories. TCIA (https://www.cancerimagingarchive.net/): collection identifier: “TCGA-BRCA.”¹⁷ TCGA (GDC, https://portal.gdc.cancer.gov/): project identifier TCGA-BRCA, with the source project’s “dbGaP study accession (TCGA: phs000178)” provided for traceability. Following multi-stage quality filtering, 840 eligible TCGA cases were retained from an initial cohort of 1097 patients through sequential exclusion of nonprimary tumors (15 cases) (Figure 1 and Supplementary Table S1), incomplete treatment-naïve status (50 cases), insufficient survival data (50 cases), male patients (12 cases), Tx-stage tumors (3 cases), and missing clinical parameters (137 cases). Concurrently, TCIA MRI datasets underwent exclusion of post-operative cases and those with suboptimal imaging quality (29/137), yielding 108 analyzable scans. Intersection of TCGA molecular data with TCIA imaging cohorts produced a final integrated dataset of 98 patients with matched radiomic-genomic profiles. Radiomic feature extraction was performed using Python’s pyradiomics library (v3.0.1), generating 2060 quantitative imaging descriptors (Supplementary Table S1). The patients were divided into 2 age groups: the “<60 years” group (ages from 18 to 59 years) and the “⩾60 years” group (ages 60 years and above). Comparative genomic analysis between PDCD1-high and PDCD1-low expression subgroups was conducted under the DESeq2 framework,¹⁸ while tumor staging followed the American Joint Committee on Cancer (AJCC) 8th Edition tumor node metastasis (TNM) Classification guidelines. Survival outcomes were evaluated through Kaplan-Meier curves with log-rank testing to assess PDCD1-associated prognostic significance.¹⁹

Figure 1.

Patient selection flowchart.

Gene expression and functional enrichment analysis

UCSC Xena (Toil Recompute): Transcriptomic expression matrices were sourced from the UCSC Toil RNA-seq Recompute data center (https://xenabrowser.net/datapages/).²⁰ Functional enrichment analysis was implemented through Gene Set Enrichment Analysis (GSEA, version 4.3.2; https://www.gsea-msigdb.org)²¹ to interrogate molecular signatures associated with PDCD1 expression. In this study, the “surv_cutpoint” function from the R package “survminer” was employed to automatically identify the most survival-significant expression threshold for PDCD1 using the maximally selected log-rank statistics (minimum P value method), which stratifies patients into subgroups with the largest survival differences. This method, extensively applied in exploratory survival research, demonstrates robust objectivity and statistical interpretability, effectively determining biomarker cutoff points with potential prognostic value. Computational analysis of the BC cohort revealed a PDCD1 expression cutoff value of 0.737600. Patients were subsequently categorized into a high-expression group (n = 387) and a low-expression group (n = 453) for further comparative analyses (Figure 1). Comprehensive clinical metadata are detailed in Supplementary Table S1. Gene Ontology (GO) annotation spanning biological processes, molecular functions, and cellular components, along with Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway analyses, were executed using the clusterProfiler package (v4.6.2)¹⁵ to deconvolve PDCD1-related biological mechanisms. Statistical significance was defined by dual significance thresholds (false discovery rate [FDR] < 0.05 and nominal P value < .05). Analytical workflows incorporated ggplot2 (v3.3.6), stats (v4.2.1), and car packages for visualization and statistical modeling.

Variance analysis of tumor mutational burden and correlation analysis

Mutation profiles in MAF format were acquired from The Cancer Genome Atlas BC (TCGA_BC) cohort. Tumor mutational burden (TMB) disparities between PDCD1-high and PDCD1-low cohorts were assessed via Wilcoxon rank-sum testing implemented through the limma package (v3.54.2). Transcriptomic matrices of BC specimens were subsequently processed through the CIBERSORTx computational platform (https://cibersortx.stanford.edu)²² to deconvolute immune cell infiltration landscapes. Bivariate correlation assessments between PDCD1 expression status (both predicted and measured) and immunological/clinical parameters were conducted using Spearman’s rank correlation coefficients, preserving nonparametric assumptions for ordinal and continuous variables.

Analysis of coherence evaluations

Volumetric tumor segmentation and lesion localization were performed using 3D Slicer (v4.10.2) with integrated multiparametric MRI sequences, including T2-weighted scanning and dynamic contrast-enhanced MRI (DCE-MRI). Imaging parameters were optimized through window width/level adjustments to identify lesion topography, supplemented by bilateral comparative analysis for detecting abnormal soft tissue masses, architectural distortions, signal intensity disparities, and contrast-enhanced heterogeneity. Following initial region of interest (ROI) annotation by a primary radiologist, a quality control subset (n = 20) was re-annotated by a secondary investigator through randomized sampling protocol (Supplementary Figure 4). Radiomic feature extraction was subsequently implemented across dual-annotated cases. Inter-rater reliability was quantified via intraclass correlation coefficient (ICC) analysis (2-way mixed-effects model) with predefined validity thresholds: excellent agreement (ICC ⩾ 0.75), moderate agreement (0.51-0.74), and poor agreement (<0.50). Features demonstrating ICC ⩾ 0.75 were retained for downstream analysis to ensure radiomic feature robustness.

Establishment of support vector machine model and logistic regression model

Robustness and reproducibility were ensured through the training-validation cohort design. Furthermore, several preprocessing steps were implemented to mitigate the impact of variations in scanning equipment, imaging protocols, and lesion size on feature extraction. N4 bias field correction: Applied to reduce signal inhomogeneity caused by magnetic field variations, enhancing image consistency; Resampling to [1,1,1] mm³ resolution: Ensured isotropic voxels, standardizing spatial resolution across datasets from different equipment/protocols; Intensity normalization (normalizeScale = 100): Reduced inter-scanner heterogeneity in pixel value distributions to improve feature comparability; Intensity discretization (binWidth = 5): Minimized noise effects on pixel values, enhancing feature calculation stability.

The dataset underwent stratified partitioning into training (n = 69) and validation (n = 29) cohorts at a 7:3 ratio, with stratification quality verified through inter-group variance analysis (Figure 1). During the preliminary feature screening phase, we identified 1694 reliable radiomic features (ICC > 0.75) by calculating the intraclass correlation coefficient (ICC) values. These features were subsequently ranked using the maximum relevance minimum redundancy (mRMR) algorithm. The mRMR method reduces multicollinearity issues by maximizing the correlation between features and the target variable while minimizing redundancy among features themselves. Subsequent recursive feature elimination yielded an optimal 5-feature subset (Supplementary Figure 1A). Predictive modeling employed dual machine learning approaches: 1) Support vector machine (SVM) regression with radial basis function kernel through the caret package (v6.0-94) for continuous gene expression prediction; 2) Logistic Regression (LR) classifier with L2 regularization via the stats package (v4.2.1) for binary gene expression categorization. Model performance was quantified across training/validation cohorts using diagnostic metrics including accuracy, specificity, sensitivity, positive/negative predictive values. Calibration fidelity was assessed through calibration plots and Hosmer-Lemeshow goodness-of-fit testing, while predictive discrimination was evaluated via Brier score. Clinical applicability was demonstrated through decision curve analysis (DCA). Analytical workflows incorporated pROC (v1.18.0), rms (v6.3-0), ResourceSelection (v0.3-6), and rmda (v1.6) packages. Comparative model performance was statistically validated using DeLong’s test for receiver operating characteristic (ROC) curve differences.

Statistical analysis

Cutoff value selection method: The probability of predicting gene expression levels output by the SVM/LR radiomics model in survmine2 is defined as the Rad_score. The LR model is computed based on the following formula: Radiomics score (Rad_score) = (Feature × corresponding coefficient [Estimate]) + Intercept value (Estimate). Univariate and multivariate Cox proportional hazards regression models were implemented to estimate hazard ratios (HRs) with corresponding 95% confidence intervals (CIs). Kaplan-Meier survival analysis with log-rank testing evaluated associations between clinical parameters and survival outcomes. Multivariable LR modeling quantified the dependence of PDCD1 expression status on clinicopathological variables. All computational workflows were executed in the R statistical computing environment (v4.1.0). Statistical significance threshold was established at α = 0.05 (2-tailed), with P values < .05 denoting significant associations.

Data availability

The datasets analyzed during the current study are available in the TCIA repository (https://www.cancerimagingarchive.net/) and TCGA repository (https://porta1.gdc.cancer.gov/).

EQUATOR network guideline

This study has followed the relevant EQUATOR Network guideline.²³

Results

Correlation between PDCD1 expression and overall survival

The 840 patients included in the survival analysis were categorized into high (n = 387) and low (n = 453) based on a PDCD1 expression cutoff value of 0.737600 (Figure 1 and Supplementary Table S1). The age (P = .047), estrogen receptor (ER) status (P < .001), histological type (P < .001), and chemotherapy treatment (P < .001) differed significantly between the 2 groups (Supplementary Table S2). In the high PDCD1 expression group, 70% of cases were ER_status positive, whereas the proportion was 84% in the low PDCD1 expression group (Supplementary Table S2). Similarly, the progesterone receptor (PR)_status positive rate was 61% in the high PDCD1 expression group and 74% in the low-expression group. The proportion of “Infiltrating Ductal Carcinoma” as the Histological_type was 73% in the high PDCD1 expression group and 83% in the low-expression group. In addition, there was a notable difference in chemotherapy administration between the groups, with 64% of the high PDCD1 expression group undergoing chemotherapy compared to 52% in the low-expression group (Supplementary Table S2).

Our analysis revealed that the gene expression level of PDCD1 in the tumor group was significantly higher than that in the normal group (P < .001) (Figure 2A). The median survival time was lower in the low PDCD1 expression group (130.87 months) when compared to the high PDCD1 expression group (219.77 months). The Kaplan-Meier survival curves showed that PDCD1 expression was significantly negatively correlated with overall survival (OS; P = .02) (Figure 2B). Patients treated with radiotherapy (P < .001), or chemotherapy (P < .001), and those below the age of 60 years (P < .001) had a significantly better OS (Figure 2C, Supplementary Figure 1B to G).

Figure 2.

Correlation between Programmed Cell Death 1 (PDCD1) expression and overall survival (OS) for specific clinical characteristics. (A) PDCD1 differential expression analysis, (B) Survival curves, (C) Correlation analysis and (D) Cox multivariate analysis.

Moreover, the univariate analysis found that high PDCD1 expression was a protective factor for OS, HR = 0.63, 95% CI 0.425-0.934, P = .021 (Figure 2D). Following multivariate adjustment, high PDCD1 expression (HR = 0.639, 95% CI 0.422-0.965, P = .033) remained a protective factor for OS (Figure 2D). Further subgroup analysis showed that the effect of PDCD1 on OS was not affected by age stratification (Supplementary Figure 2).

Association between PDCD1 expression levels and clinical characteristics

Correlation analysis showed that tumor patients receiving chemotherapy were more inclined to have high PDCD1 expression (P < .001), whereas estrogen receptor (ER) and progesterone receptor (PR) levels in patients with breast cancer (BC) were significantly negatively correlated with PDCD1 expression (P < .001) (Figure 2C). However, chemotherapy (P < .001) was significantly negatively correlated with ER status (P < .001) and PR status (P < .001) (Figure 2C).

Correlation between PDCD1 expression levels and immune cell infiltration

The PDCD1 expression levels showed a significant positive correlation with most immune cell infiltrates, such as T cells CD8 (P < .001), T cells regulatory (Treg) (P < .001), T cells follicular helper (P < .001), memory activated T cells CD4 (P < .001), memory resting T cells CD4 (P < .001), and activated natural killer (NK) cells (P < .001) (Figure 3A). PDCD1 was also significantly negatively correlated with eosinophils (P < .001), activated mast cells (P < .001), resting NK cells (P < .001), and memory B cells (P < .001) (Figure 3A). The correlation highlighted the essential role of the tumor microenvironment in determining the effectiveness of treatments and the survival prospects for patients. Furthermore, the expression of PDCD1 was found to be significantly and positively correlated with the level of immune cell infiltration.

Figure 3

Differential expression and functional analysis of PDCD1. (A) Correlation analysis between PDCD1 and immune cell infiltration, (B) Analysis of PDCD1 and tumor mutational burden (TMB); Top 20 significantly enriched pathways based on (C) Hallmark enrichment analysis and (D) Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analysis.

Identification of differentially expressed genes and enrichment analysis

The high PDCD1 expression group had significantly higher levels of TMB than the PDCD1 low-expression group (P < .001) (Figure 3B). These results suggest that patients with high TMB may respond better to immune-targeted therapy. The Kyoto Encyclopedia of Genes and Genomes (KEGG) differentially expressed genes analysis showed that the T cell receptor (TCR) signaling pathway, B cell receptor (BCR) signaling pathway, toll-like receptor (TLR) signaling pathway, antigen processing and presentation, NK cell-mediated cytotoxicity, hematopoietic cell lineage, allograft rejection, the intestinal immune network for immunoglobulin A (IgA) production, chemokine signaling pathway, and the cytokine-cytokine receptor were significantly enriched in the high PDCD1 expression group (Figure 3C).

In the Hallmark gene set, the differentially expressed genes were significantly enriched for the apoptosis pathway, Interleukin 2-Signal Transducer and Activator of Transcription 5 (IL-2-STAT5) signaling, Kirsten rat sarcoma viral oncogene homolog (KRAS) signaling, Interleukin 6-Janus Kinase-Signal Transducer and Activator of Transcription 3 (IL-6-JAK-STAT3) signaling, Tumor Necrosis Factor alpha (TNF-α) signaling via Nuclear Factor kappa-light-chain-enhancer of activated B cells (NF-κB), inflammatory response, Interferon alpha (IFNα) response, allograft rejection, and Interferon gamma (IFNγ) response pathways (Figure 3D).

High clinical utility and accuracy of the PDCD1 expression level MRI radiomics models

The patients were randomly divided into training (n = 69) and validation (n = 29) groups. There was no statistically significant difference in the baseline characteristics between the 2 groups (Supplementary Table S3). The Recursive Feature Elimination (RFE), and minimum Redundancy Maximum Relevance (mRMR) screening methods revealed 5 radiomics features predictive of PDCD1 expression (Supplementary Figure 1A). These features were used to train the SVM (Supplementary Figure 3A) and LR learning models (Supplementary Figure 3B).

The area under the curve (AUC) for the SVM model was 0.808 for the training dataset (Figure 4A) and 0.729 for the validation dataset (Figure 4B). The LR model had an AUC of 0.819 in the training dataset (Figure 4C) and 0.767 in the validation dataset (Figure 4D). These results indicated that both models had good prediction performance on both the training and validation datasets. The calibration curves and Hosmer-Lemeshow goodness-of-fit tests showed a good agreement between the predicted PDCD1 levels and the true values for both the SVM and LR models (P > .05) (Figure 4I to J). The Decision Curve Analysis (DCA) showed that both models had high clinical utility (Figure 4E to H). The high PDCD1 expression group had a significantly higher radiomics score than the low PDCD1 expression group on the training and validation datasets of both models (P < .05) (Figure 4M to P). Furthermore, the radiomics score showed a significant positive correlation with the PDCD1 expression levels, immune cell infiltration, including CD8 and Treg, the B and T Lymphocyte Attenuator (BTLA), Cytotoxic T-lymphocyte-associated protein 4 (CTLA4), Inducible T-Cell Costimulator (ICOS), Lymphocyte-Activation Gene 3 (LAG3), and T-Cell Immunoreceptor with Ig and ITIM domains (TIGIT) (P < .001, R > 0.54) (Figure 5).

Figure 4.

Evaluation of model efficacy in the training and validation sets. (A) and (B) receiver operating characteristic (ROC) curves analysis for the training and validation datasets for the Support Vector Machine (SVM) model. (C) and (D) ROC curves analysis for the training and validation sets for the Logistic Regression (LR) model. (E) and (F) Decision Curve Analysis (DCA) for the training and validation datasets for the SVM model. (G) and (H) DCA for the training and validation sets for the LR model. (I) and (J) calibration curves and Hosmer-Lemeshow goodness-of-fit test for the training and validation datasets for the SVM model. (K) and (L) calibration curves and Hosmer-Lemeshow goodness-of-fit test for the training and validation datasets for the LR model. (M) and (N) the predicted PDCD1 expression level (Rad_score) for the training and validation datasets for the SVM model. (O) and (P) Rad_score for the training and validation datasets for the LR model.

Figure 5

Correlation between the radiomics score, PDCD1, and immune gene predicted by the logistic regression (LR) radiomics model.

Discussion

The early characterization of BC could facilitate the provision of effective early interventions necessary to improve outcomes. Current methods used to characterize BC require invasive expensive tissue sampling and analysis. Therefore, in this study, we investigated the prognostic significance of PDCD1 expression levels in BC. In addition, we also made use of MRI radiomics features to develop a model that could be used to facilitate the assessment of PDCD1 expression levels through noninvasive imaging.

Our findings revealed that elevated PDCD1 levels were significantly correlated with several key signaling pathways, including the IL2/STAT5, KRAS, TNF-α-NF-κB, T cell receptor, and B cell receptor signaling pathways. This correlation suggests that PDCD1 may play a central role in the complex network of BC pathophysiology. For instance, type I regulatory T cells (Tr1 cells) are a subpopulation of CD4 T cells that play an important role in the negative regulation of immune response and the maintenance of the body’s immune response, while the IL-2/STAT is an important signaling pathway controlling the differentiation and proliferation of the CD4 T-cell subpopulation.²⁴

Transcriptional programs regulated by the NF-κB are essential for the development and maintenance of the immune system, skeletal system, and epithelium.²⁵ The T lymphocytes are also an important component of the immune system.²⁶ The B cell receptor (BCR) signaling pathway is an important signaling pathway for the survival of B cells.²⁷ By uncovering the intricate interplay between PDCD1 expression and these critical signaling pathways, our study provides valuable insights into the potential impact of PDCD1 in BC biology and opens avenues for further research into the use of targeted therapies and precision medicine for BC.

These above findings indicate that PDCD1 expression may provide important information on the immune response and the selection of appropriate targeted therapy. Therefore, we decided to explore whether MRI could provide a noninvasive tool to measure PDCD1 expression. Our results indicate that the radiomics score positively correlated with the PDCD1 and the expression levels of several immune genes, including BTLA, CTLA4, ICOS, LAG3, and TIGIT.²⁸ BTLA is expressed in tumor-infiltrating lymphocytes (TILs) and is often associated with impaired anti-tumor immune response. Studies have shown that up-regulation of BTLA expression in gallbladder cancer has a role in suppressing anti-cancer immunity.²⁹ BTLA has an inhibitory function in controlling T-cell activation in vivo. CTLA4 is a protein receptor that acts as an immune checkpoint inhibitor.³⁰ CTLA4 is constitutively expressed in regulatory T cells but is only up-regulated in regular T cells after activation. Overactivation of CTLA4 is common in cancer cells.

This study has several limitations that have to be acknowledged. First, we cannot completely avoid potential selection bias during sample screening, particularly when excluded cases demonstrate systematic differences in clinical or biological characteristics from included cohorts. These differences may compromise model generalizability and stability. Therefore, we recommend future validation in larger, multicenter prospective cohorts to enhance robustness and clinical utility. Second, this study trained and validated models exclusively using TCGA and TCIA data, which constitutes a major limitation as we did not conduct independent external validation. This absence may impair generalization capability and the reliability of clinical deployment. We, therefore, consider validation in larger multicenter prospective cohorts necessary to strengthen robustness and clinical applicability. Future research should explore integrating imaging features with key clinical variables to develop multimodal predictive models, potentially enhancing predictive performance and enabling comprehensive stability verification.

Furthermore, this study has several clinical implementation barriers. First, the effective integration of radiomic tools into existing clinical workflows necessitates enhanced technical platforms, accelerated data processing capabilities, and comprehensive technical staff training. Second, standardization challenges arise from heterogeneities in imaging protocols, scanning equipment, and acquisition parameters across institutions, impacting feature extraction and model application reliability. Finally, multisite, multi-scanner validation is essential to ensure model robustness and reproducibility. Hence, advancing clinical applicability requires optimizing preprocessing techniques and feature harmonization protocols, with subsequent validation in large-scale, multicenter cohorts to amplify generalizability. In practice, the radiomics score is intended to operate in parallel with clinical variables as an independent molecular-phenotype proxy, enabling complementary risk stratification rather than replacing established clinical factors. In addition, while most patients acknowledge the potential benefits of machine learning and artificial intelligence (AI) in oncology, concerns remain regarding diagnostic accuracy, data privacy, and medical costs. In addition, educational level is associated with acceptance. Therefore, when promoting the clinical adoption of AI imaging tools, patient education and transparent ethical and data compliance strategies should be implemented concurrently to enhance acceptability and practical effectiveness.³¹

Conclusion

High PDCD1 expression levels were associated with improved OS. PDCD1 expression levels also showed a significant positive correlation with immune cell infiltration. The MRI radiomics model developed in our study could potentially serve as a noninvasive tool for the early prediction of PDCD1 expression levels in BC and thus facilitate the delivery of effective early interventions.

Supplemental Material

sj-pdf-1-onc-10.1177_11795549251399383 – Supplemental material for Machine Learning–Based Enhanced MRI Radiomics for PDCD1 Prognostication and Expression Prediction in Breast Cancer

Supplemental material, sj-pdf-1-onc-10.1177_11795549251399383 for Machine Learning–Based Enhanced MRI Radiomics for PDCD1 Prognostication and Expression Prediction in Breast Cancer by Yingying Gao, Zihan Li, Ziyun Li and Xueyan Gao in Clinical Medicine Insights: Oncology

Supplemental Material

sj-xlsx-2-onc-10.1177_11795549251399383 – Supplemental material for Machine Learning–Based Enhanced MRI Radiomics for PDCD1 Prognostication and Expression Prediction in Breast Cancer

Supplemental material, sj-xlsx-2-onc-10.1177_11795549251399383 for Machine Learning–Based Enhanced MRI Radiomics for PDCD1 Prognostication and Expression Prediction in Breast Cancer by Yingying Gao, Zihan Li, Ziyun Li and Xueyan Gao in Clinical Medicine Insights: Oncology

Supplemental Material

sj-xlsx-3-onc-10.1177_11795549251399383 – Supplemental material for Machine Learning–Based Enhanced MRI Radiomics for PDCD1 Prognostication and Expression Prediction in Breast Cancer

Supplemental material, sj-xlsx-3-onc-10.1177_11795549251399383 for Machine Learning–Based Enhanced MRI Radiomics for PDCD1 Prognostication and Expression Prediction in Breast Cancer by Yingying Gao, Zihan Li, Ziyun Li and Xueyan Gao in Clinical Medicine Insights: Oncology

Supplemental Material

sj-xlsx-4-onc-10.1177_11795549251399383 – Supplemental material for Machine Learning–Based Enhanced MRI Radiomics for PDCD1 Prognostication and Expression Prediction in Breast Cancer

Supplemental material, sj-xlsx-4-onc-10.1177_11795549251399383 for Machine Learning–Based Enhanced MRI Radiomics for PDCD1 Prognostication and Expression Prediction in Breast Cancer by Yingying Gao, Zihan Li, Ziyun Li and Xueyan Gao in Clinical Medicine Insights: Oncology

Footnotes

Acknowledgements

We would like to thank TopEdit () for the English language editing of this manuscript.

ORCID iD

Yingying Gao

Ethical considerations

All imaging and clinical data used in this study were obtained from the publicly available TCIA/TCGA database (). These data were fully de-identified and publicly accessible, and their use complied with the database’s data usage policies. No personal identifiers were accessible to the authors at any time.

Consent to participate

Not Applicable.

Consent for publication

Not Applicable.

Author contribution

YG made a substantial contribution to the concept and design of the work, and analyzed the data and drafted and revised the article, and approved the final version to be published. Zihan L analyzed the data and approved the final version to be published. Ziyun L made a substantial contribution to the concept and design of the work, drafted the article, and approved the final version to be published. XG designed the work and drafted and revised the article, and approved the final version to be published.

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Supplemental material

Supplemental material for this article is available online.

References

Harbeck

Penault-Llorca

Cortes

, et al. Breast cancer. Nat Rev Dis Primers. 2019;5:66. doi:10.1038/s41572-019-0111-2

Nolan

Lindeman

Visvader

JE.

Deciphering breast cancer: from biology to the clinic. Cell. 2023;186:1708-1728. doi:10.1016/j.cell.2023.01.040

Yau

Osdoit

van der Noordaa

, et al. Residual cancer burden after neoadjuvant chemotherapy and long-term survival outcomes in breast cancer: a multicentre pooled analysis of 5161 patients. Lancet Oncol. 2022;23:149-160. doi:10.1016/s1470-2045(21)00589-1

Rakha

Pareja

FG.

New advances in molecular breast cancer pathology. Semin Cancer Biol. 2021;72:102-113. doi:10.1016/j.semcancer.2020.03.014

Miao

Wang

, et al. Prognostic value and immunological role of PDCD1 gene in pan-cancer. Int Immunopharmacol. 2020;89:107080. doi:10.1016/j.intimp.2020.107080

Wartewig

Daniels

Schulz

, et al. PD-1 instructs a tumor-suppressive metabolic program that restricts glycolysis and restrains AP-1 activity in T cell lymphoma. Nat Cancer. 2023;4:1508-1525. doi:10.1038/s43018-023-00635-7

Sharpe

Pauken

KE.

The diverse functions of the PD1 inhibitory pathway. Nat Rev Immunol. 2018;18:153-167. doi:10.1038/nri.2017.108

Kopans

Meyer

Sadowsky

Breast imaging. N Engl J Med. 1984;310:960-967. doi:10.1056/nejm198404123101506

Lehman

Schnall

MD.

Imaging in breast cancer: magnetic resonance imaging. Breast Cancer Research. 2005;7:215. doi:10.1186/bcr1309

10.

Conti

Duggento

Indovina

Guerrisi

Toschi

Radiomics in breast cancer classification and prediction. Semin Cancer Biol. 2021;72:238-250. doi:10.1016/j.semcancer.2020.04.002

11.

Duggento

Conti

Mauriello

Guerrisi

Toschi

Deep computational pathology in breast cancer. Semin Cancer Biol. 2021;72:226-237. doi:10.1016/j.semcancer.2020.08.006

12.

Schabath

Gillies

RJ.

Images are data: challenges and opportunities in the clinical translation of radiomics. Cancer Res. 2022;82:2066-2068. doi:10.1158/0008-5472.Can-22-1183

13.

Gao

Jiang

Yue

, et al. Radiomic model to predict the expression of PD-1 and overall survival of patients with ovarian cancer. Int Immunopharmacol. 2022;113:109335. doi:10.1016/j.intimp.2022.109335

14.

Wang

The application of radiomics in breast MRI: a review. Technol Cancer Res Treat. 2020;19:1533033820916191. doi:10.1177/1533033820916191

15.

Wang

Han

QY.

ClusterProfiler: an R package for comparing biological themes among gene clusters. OMICS. 2012;16:284-287. doi:10.1089/omi.2011.0118

16.

Campana

Gandomkar

Giannotti

Reed

The use of radiomics in magnetic resonance imaging for the pre-treatment characterisation of breast cancers: a scoping review. J Med Radiat Sci. 2023;70:462-478. doi:10.1002/jmrs.709

17.

Lingle

Erickson

Zuley

, et al. The cancer genome atlas breast invasive carcinoma collection (TCGA-BRCA) (version 3) [Data set]. The Cancer Imaging Archive, 2016. https://doi.org/10.7937/K9/TCIA.2016.AB2NAZRP

18.

Love

Huber

Anders

Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2. Genome Biol. 2014;15:550. doi:10.1186/s13059-014-0550-8

19.

Liu

Lichtenberg

Hoadley

, et al. An integrated TCGA pan-cancer clinical data resource to drive high-quality survival outcome analytics. Cell. 2018;173:400-416.e11. doi:10.1016/j.cell.2018.02.052

20.

Vivian

Rao

Nothaft

, et al. Toil enables reproducible, open source, big biomedical data analyses. Nat Biotechnol. 2017;35:314-316. doi:10.1038/nbt.3772

21.

Subramanian

Tamayo

Mootha

, et al. Gene set enrichment analysis: a knowledge-based approach for interpreting genome-wide expression profiles. Proc Natl Acad Sci U S A. 2005;102:15545-15550. doi:10.1073/pnas.0506580102

22.

Newman

Steen

Liu

, et al. Determining cell type abundance and expression from bulk tissues with digital cytometry. Nat Biotechnol. 2019;37:773-782. doi:10.1038/s41587-019-0114-2

23.

Bossuyt

Reitsma

Bruns

, et al. STARD 2015: an updated list of essential items for reporting diagnostic accuracy studies. BMJ. 2015;351:h5527. doi:10.1136/bmj.h5527

24.

Polhill

Zhang

, et al. IL-2/IL-2Ab complexes induce regulatory T cell expansion and protect against proteinuric CKD. J Am Soc Nephrol. 2012;23:1303-1308. doi:10.1681/asn.2011111130

25.

Ilchovska

Barrow

DM.

An overview of the NF-kB mechanism of pathophysiology in rheumatoid arthritis, investigation of the NF-kB ligand RANKL and related nutritional interventions. Autoimmun Rev. 2021;20:102741. doi:10.1016/j.autrev.2020.102741

26.

Shah

Al-Haidari

Sun

Kazi

JU.

T cell receptor (TCR) signaling in health and disease. Signal Transduct Target Ther. 2021;6:412. doi:10.1038/s41392-021-00823-w

27.

Berning

Lenz

Targeting B-cell receptor and PI3K signaling in diffuse large B-cell lymphoma. Blood. 2021;138:1110-1119. doi:10.1182/blood.2020006784

28.

Xie

Chen

Tang

, et al. A Necroptosis-related prognostic model of uveal melanoma was constructed by single-cell sequencing analysis and weighted co-expression network analysis based on public databases. Front Immunol. 2022;13:847624. doi:10.3389/fimmu.2022.847624

29.

Oguro

Ino

Shimada

, et al. Clinical significance of tumor-infiltrating immune cells focusing on BTLA and Cbl-b in patients with gallbladder cancer. Cancer Sci. 2015;106:1750-1760. doi:10.1111/cas.12825

30.

Pai

Simons

, et al. Tumor-conditional anti-CTLA4 uncouples antitumor efficacy from immunotherapy-related toxicity. J Clin Invest. 2019;129:349-363. doi:10.1172/jci123391

31.

Erul

Aktekin

Danışman

, et al. Perceptions, attitudes, and concerns on artificial intelligence applications in patients with cancer. Cancer Control. 2025;32:10732748251343245. doi:10.1177/10732748251343245

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

4.00 MB

0.01 MB

0.07 MB