Historically, medical imaging has been a qualitative or semi-quantitative modality. It is difficult to quantify what can be seen in an image, and to turn it into valuable predictive outcomes. As a result of advances in both computational hardware and machine learning algorithms, computers are making great strides in obtaining quantitative information from imaging and correlating it with outcomes. Radiomics, in its two forms “handcrafted and deep,” is an emerging field that translates medical images into quantitative data to yield biological information and enable radiologic phenotypic profiling for diagnosis, theragnosis, decision support, and monitoring. Handcrafted radiomics is a multistage process in which features based on shape, pixel intensities, and texture are extracted from radiographs. Within this review, we describe the steps: starting with quantitative imaging data, how it can be extracted, how to correlate it with clinical and biological outcomes, resulting in models that can be used to make predictions, such as survival, or for detection and classification used in diagnostics. The application of deep learning, the second arm of radiomics, and its place in the radiomics workflow is discussed, along with its advantages and disadvantages. To better illustrate the technologies being used, we provide real-world clinical applications of radiomics in oncology, showcasing research on the applications of radiomics, as well as covering its limitations and its future direction.

Stability analysis remains a fundamental step in developing a successful imaging biomarker to personalize oncological strategies. This study proposes an in silico contour generation method for simulating segmentation variations to identify stable radiomic features. Ground-truth annotation provided for the whole prostate gland on the multi-parametric MRI sequences (T2w, ADC, and SUB-DCE) were perturbed to mimic segmentation differences observed among human annotators. In total, we generated 15 synthetic contours for a given image-segmentation pair. One thousand two hundred twenty-four unfiltered/filtered radiomic features were extracted applying Pyradiomics, followed by stability assessment using ICC(1,1). Stable features identified in the internal population were then compared with an external population to discover and report robust features. Finally, we also investigated the impact of a wide range of filtering strategies on the stability of features. The percentage of unfiltered (filtered) features that remained robust subjected to segmentation variations were T2w—36% (81%), ADC—36% (94%), and SUB—43% (93%). Our findings suggest that segmentation variations can significantly impact radiomic feature stability but can be mitigated by including pre-filtering strategies as part of the feature extraction pipeline.

Objectives: This study aimed to assess whether pharmacokinetic parameters derived from DCE-MRI can stratify Programmed Death-Ligand 1 (PD-L1) expression in NSCLC. The secondary aim was to identify a suitable pharmacokinetic model configuration for anisotropic temporally-spaced DCE-MRI sequences, considering Tofts variants, population-averaged arterial input functions (AIF), and bolus arrival time (BAT) estimation methods. Materials and methods: From April 2021 to May 2023, patients with locally advanced non-small cell lung cancer (NSCLC) were prospectively enrolled. Tumors were categorized based on: PD-L1 absence/presence (threshold 1%) and hyperexpression/hypoexpression (threshold 50%). Pharmacokinetic parameters were extracted using several candidate configurations; fit quality was evaluated using coefficient of determination (R²). Mann–Whitney U-test and ROC-AUC were used to assess correlation with PD-L1 for the best-fit configuration. Results: Thirty-eight patients (mean age 68 ± 9 years, 28 men) were included. PD-L1 expression was present in 25 patients (66%) and absent in 13 (34%). PD-L1 was hyperexpressed in 13 (34%) patients and hypoexpressed in 25 (66%). Voxel-wise pharmacokinetic parameters were extracted using the best-fit configuration—extended Tofts model (ETM) with Georgiou AIF and Peak-Gradient (PG) BAT estimation (R ² = 0.79). K^trans median (0.25 vs. 0.12 min⁻¹, p = 0.02), K^trans standard deviation (0.32 vs. 0.23 min⁻¹, p = 0.01) and K_ep median (1.09 vs. 0.59 min⁻¹, p = 0.02) were significantly higher in PD-L1 < 50% group (ROC-AUC 0.71–0.76). Conclusion: DCE-MRI pharmacokinetic parameters could stratify PD-L1 hypo/hyperexpression in NSCLC. The ETM with PG BAT estimation method and Georgiou AIF was the best-performing pharmacokinetic configuration. Key Points: Question Could Dynamic Contrast-Enhanced (DCE) MRI offer a safe and non-invasive way to assess Programmed Death-Ligand 1 (PD-L1) expression? Findings Quantitative DCE-MRI parameters K ^trans (the volume transfer rate) and K_ep (the efflux rate constant) show potential for distinguishing PD-L1 hyperexpression from hypoexpression. Clinical relevance Preliminary results suggest that DCE-MRI could be a safe method to stratify PD-L1 hypo/hyperexpression in non-small cell lung cancer, potentially optimizing treatment decisions, given the high cost of immunotherapy. Graphical

Introduction: Acute appendicitis (AA) is a common cause of abdominal pain that can lead to complications like perforation and intra-abdominal abscesses, increasing morbidity and mortality, often requiring emergency surgery. Nevertheless, appendectomy is performed in up to 95% of uncomplicated cases, while complications like perforation and intra-abdominal abscesses increase morbidity and mortality. The current study compares the accuracy of GPT-4.5, DeepSeek R1, and machine learning in assisting with surgical decision-making for patients presenting with lower abdominal pain at the Emergency Department. Methods: In this multicenter retrospective study, 63 histopathologically confirmed appendicitis patients and 50 control patients with right abdominal pain presenting at the Emergency Department at two German hospitals between October 2022 and October 2023 were included. Using each patient’s clinical, laboratory, and radiological findings, DeepSeek (with and without Retrieval-Augmented Generation using 2020 Jerusalem guidelines) was compared in terms of accuracy with GPT-4.5 and a random forest-based machine-learning model, with a board-certified surgeon (reference standard) to determine the optimal treatment approach (laparoscopic exploration/appendectomy versus conservative antibiotic therapy). Results: Accuracy of agreement with board-certified surgeons in the decision-making of appendectomy versus conservative therapy increased non-significantly from 80.5% to 83.2% with DeepSeek and from 70.8 to 76.1% when GPT-4.5 was provided with the World Journal of Emergency Surgery 2020 Jerusalem guidelines on the diagnosis and treatment of acute appendicitis. The estimated machine-learning model training accuracy was 84.3%, while the validation accuracy for the model was 85.0%. Discussion: GPT-4.5 and DeepSeek R1, as well as the machine-learning model, demonstrate promise in aiding surgical decision-making for appendicitis, particularly in resource-constrained settings. Ongoing training and validation are required to optimize the performance of such models.

Background/Objectives: In skull-base proton therapy, T2-weighted brain magnetic resonance (MR) imaging may be acquired with or without fat saturation (FS). The present study investigated the impact of this protocol variation on radiomics feature stability and brain radionecrosis (BRN) prediction. Methods: Paired T2-weighted FS and non-FS follow-up MR scans of proton-treated skull-base chordoma patients (n = 52) were used to assess feature stability. For BRN prediction (CTCAEv5 grade ≥ 1), baseline planning scans of chordoma and chondrosarcoma patients (n = 80) with mixed FS protocols were used. Following automated brain tissue segmentation, 1911 radiomics features were extracted from cerebrospinal fluid, gray, and/or white matter using PyRadiomics (v3.1.0). Feature stability was quantified using Lin’s concordance correlation coefficient (CCC). Several MR image- and feature-level processing configurations were explored, and the stability of combined gray and white matter features served as the reference for selecting the optimal configuration. Features were stratified based on CCC thresholds and were subjected to univariable feature selection using a Mann–Whitney U-test. Logistic regression was used for predictive modeling and its performance was measured using micro-averaged AUC within a repeated stratified cross-validation framework. Results: Radiomics features were highly sensitive to FS variations (median CCC range: 0.21–0.58). Combat harmonization without image-level processing was selected as the optimal configuration, under which 30% of features achieved CCC ≥ 0.70, and 10.7% demonstrated a CCC ≥ 0.85. Restricting modeling to a highly stable subset eliminated nearly 90% of the baseline features while significantly improving predictive performance (ΔAUC = 0.05, p < 0.001). Conclusions: A subset of radiomics features robust to FS variations was identified. Use of these stable features may mitigate the impact of protocol-induced variability in mixed-FS T2-weighted MR data while simultaneously improving BRN prediction performance. These findings are preliminary and should be interpreted cautiously, given the pilot nature of the study.

Objective. Lung cancer remains the leading cause of cancer-related mortality worldwide, with Non-Small Cell Lung Cancer (NSCLC) accounting for approximately 85% of all cases. Programmed cell Death Ligand-1 (PD-L1) is a well-established biomarker that guides immunotherapy in advanced-stage NSCLC, currently evaluated via invasive biopsy procedures. This study aims to develop and validate a non-invasive pipeline for stratifying PD-L1 expression using quantitative analysis of IVIM parameter maps—diffusion (D), pseudo-diffusion (D*), perfusion fraction (pf)—and T1-VIBE MRI acquisitions. Approach. MRI data from 43 NSCLC patients were analysed and labelled as PD-L1 positive (≥1%) or negative (<1%) based on immunohistochemistry exam. After pre-processing, 1,171 radiomic features and 512 deep learning features were obtained. Three feature sets (radiomic, deep learning, and fusion) were tested with Logistic Regression, Random Forest, and XGBoost. Four discriminative features were selected using the Mann–Whitney U-test, and model performance was primarily assessed using the area under the receiver operating characteristic curve (AUC). Robustness was ensured through repeated stratified 5-fold cross-validation, bootstrap-derived confidence intervals, and permutation test. Main Results. Logistic Regression generally demonstrated the highest classification performance, with AUC values ranging from 0.78 to 0.92 across all feature sets. Fusion models outperformed or matched the performance of the best standalone radiomics or deep learning model. Among multisequence MRI, the IVIM-D fusion features yielded the best performance with an AUC of 0.92, followed by IVIM-D* radiomic features that showed a similar AUC of 0.91. For IVIM-pf and T1-VIBE derived features, the fusion model yielded the best AUC values of 0.87 and 0.90, respectively. Significance. The obtained results highlight the potential of a combined radiomic-deep learning approach to effectively detect PD-L1 expression from MRI acquisitions, paving the way for a non-invasive PD-L1 evaluation procedure.