Foundation models in deep learning are characterized by a single large-scale model trained on vast amounts of data serving as the foundation for various downstream tasks. Foundation models are generally trained using self-supervised learning and excel in reducing the demand for training samples in downstream applications. This is especially important in medicine, where large labelled datasets are often scarce. Here, we developed a foundation model for cancer imaging biomarker discovery by training a convolutional encoder through self-supervised learning using a comprehensive dataset of 11,467 radiographic lesions. The foundation model was evaluated in distinct and clinically relevant applications of cancer imaging-based biomarkers. We found that it facilitated better and more efficient learning of imaging biomarkers and yielded task-specific models that significantly outperformed conventional supervised and other state-of-the-art pretrained implementations on downstream tasks, especially when training dataset sizes were very limited. Furthermore, the foundation model was more stable to input variations and showed strong associations with underlying biology. Our results demonstrate the tremendous potential of foundation models in discovering new imaging biomarkers that may extend to other clinical use cases and can accelerate the widespread translation of imaging biomarkers into clinical settings.

Background: Prostate cancer is a major health concern in aging men. Paralleling an aging society, prostate cancer prevalence increases emphasizing the need for efficient diagnostic algorithms. Methods: Retrospectively, 106 prostate tissue samples from 48 patients (mean age, $$66\pm 6.6$$66 ± 6.6 years) were included in the study. Patients suffered from prostate cancer (n = 38) or benign prostatic hyperplasia (n = 10) and were treated with radical prostatectomy or Holmium laser enucleation of the prostate, respectively. We constructed tissue microarrays (TMAs) comprising representative malignant (n = 38) and benign (n = 68) tissue cores. TMAs were processed to histological slides, stained, digitized and assessed for the applicability of machine learning strategies and open–source tools in diagnosis of prostate cancer. We applied the software QuPath to extract features for shape, stain intensity, and texture of TMA cores for three stainings, H&E, ERG, and PIN-4. Three machine learning algorithms, neural network (NN), support vector machines (SVM), and random forest (RF), were trained and cross-validated with 100 Monte Carlo random splits into 70% training set and 30% test set. We determined AUC values for single color channels, with and without optimization of hyperparameters by exhaustive grid search. We applied recursive feature elimination to feature sets of multiple color transforms. Results: Mean AUC was above 0.80. PIN-4 stainings yielded higher AUC than H&E and ERG. For PIN-4 with the color transform saturation, NN, RF, and SVM revealed AUC of $$0.93\pm 0.04$$0.93 ± 0.04 , $$0.91\pm 0.06$$0.91 ± 0.06 , and $$0.92\pm 0.05$$0.92 ± 0.05 , respectively. Optimization of hyperparameters improved the AUC only slightly by 0.01. For H&E, feature selection resulted in no increase of AUC but to an increase of 0.02–0.06 for ERG and PIN-4. Conclusions: Automated pipelines may be able to discriminate with high accuracy between malignant and benign tissue. We found PIN-4 staining best suited for classification. Further bioinformatic analysis of larger data sets would be crucial to evaluate the reliability of automated classification methods for clinical practice and to evaluate potential discrimination of aggressiveness of cancer to pave the way to automatic precision medicine.

Objectives: To analyze the performance of radiological assessment categories and quantitative computational analysis of apparent diffusion coefficient (ADC) maps using variant machine learning algorithms to differentiate clinically significant versus insignificant prostate cancer (PCa). Methods: Retrospectively, 73 patients were included in the study. The patients (mean age, 66.3 ± 7.6 years) were examined with multiparametric MRI (mpMRI) prior to radical prostatectomy (n = 33) or targeted biopsy (n = 40). The index lesion was annotated in MRI ADC and the equivalent histologic slides according to the highest Gleason Grade Group (GrG). Volumes of interest (VOIs) were determined for each lesion and normal-appearing peripheral zone. VOIs were processed by radiomic analysis. For the classification of lesions according to their clinical significance (GrG ≥ 3), principal component (PC) analysis, univariate analysis (UA) with consecutive support vector machines, neural networks, and random forest analysis were performed. Results: PC analysis discriminated between benign and malignant prostate tissue. PC evaluation yielded no stratification of PCa lesions according to their clinical significance, but UA revealed differences in clinical assessment categories and radiomic features. We trained three classification models with fifteen feature subsets. We identified a subset of shape features which improved the diagnostic accuracy of the clinical assessment categories (maximum increase in diagnostic accuracy ΔAUC = + 0.05, p < 0.001) while also identifying combinations of features and models which reduced overall accuracy. Conclusions: The impact of radiomic features to differentiate PCa lesions according to their clinical significance remains controversial. It depends on feature selection and the employed machine learning algorithms. It can result in improvement or reduction of diagnostic performance. Key Points: • Quantitative imaging features differ between normal and malignant tissue of the peripheral zone in prostate cancer. • Radiomic feature analysis of clinical routine multiparametric MRI has the potential to improve the stratification of clinically significant versus insignificant prostate cancer lesions in the peripheral zone. • Certain combinations of standard multiparametric MRI reporting and assessment categories with feature subsets and machine learning algorithms reduced the diagnostic performance over standard clinical assessment categories alone.

Although immunotherapy has revolutionized cancer treatment, many patients still experience limited benefit, highlighting the urgent need for improved biomarkers¹. Although immunotherapy is founded on unleashing T cells², most existing biomarkers remain tumour-centric and mainly overlook host immune competence. The thymus is a key immune organ that is crucial for T cell maturation, and we hypothesized that thymic functionality is associated with immunotherapy outcomes³. Here we show that thymic health, a radiographic measure of thymic functionality, is strongly associated with immunotherapy outcomes across several cancer types. Using a deep-learning framework applied to routine computed tomography images, we quantified thymic health in a pan-cancer cohort of 3,476 patients receiving immune checkpoint inhibitors. In patients with non-small cell lung cancer, higher thymic health was associated with reduced risks of progression and all-cause mortality. These associations remained significant across clinically relevant levels of programmed death ligand 1 (PD-L1) and tumour mutation burden. In the prospective TRACERx lung cancer study, thymic health was positively associated with T cell receptor diversity and T cell receptor excision circles, and correlated with immune-system signalling pathways, supporting radiographic thymic health as a proxy for thymic activity and adaptive immune competence. Analysis across patients with melanoma, breast cancer or renal cancer demonstrated pan-cancer relevance. Together, these findings identify thymic health as a previously unrecognized, tumour-agnostic determinant of immunotherapy efficacy, with potential implications for patient stratification, treatment timing and the development of immune-rejuvenating strategies in precision immuno-oncology.

The thymus is essential for establishing T cell diversity early in life, but undergoes profound involution with age and has therefore traditionally been regarded as largely nonfunctional in adults^1,2. Here we propose that preserving thymic functionality is integral to adult health and longevity. We developed a deep learning framework to quantify thymic health from routine radiographic images and evaluated its association with longevity and risk of major age-associated diseases in two large prospective cohorts of asymptomatic adults: the National Lung Screening Trial (n = 25,031) and the Framingham Heart Study (n = 2,581). In both cohorts, thymic health varied markedly across the population. In the National Lung Screening Trial, higher thymic health was consistently associated with lower all-cause mortality, reduced lung cancer incidence and lower cardiovascular mortality over 12 years of follow-up after adjustment for age, sex, smoking and comorbidities. In the independent Framingham Heart Study cohort, higher thymic health was significantly associated with reduced cardiovascular mortality, independent of age, sex and smoking. Thymic health was further linked to systemic inflammation and metabolic dysregulation, and associated with modifiable lifestyle factors including smoking, obesity and physical activity. Together, these findings reposition the thymus as a central regulator of immune-mediated ageing and disease susceptibility in adulthood, highlighting its potential as a target for preventive and regenerative strategies to promote healthy ageing and longevity.

Seizures are among the most common symptoms of meningioma. This retrospective study sought to identify risk factors for early and late seizures in meningioma patients and to evaluate a modified STAMPE2 score. In 556 patients who underwent meningioma surgery, we correlated different risk factors with the occurrence of postoperative seizures. A modified STAMPE2 score was applied. Risk factors for preoperative seizures were edema (p = 0.039) and temporal location (p = 0.038). For postoperative seizures preoperative tumor size (p < 0.001), sensomotory deficit (p = 0.004) and sphenoid wing location (p = 0.032) were independent risk factors. In terms of postoperative status epilepticus; sphenoid wing location (p = 0.022), tumor volume (p = 0.045) and preoperative seizures (p < 0.001) were independent risk factors. Postoperative seizures lead to a KPS deterioration and thus an impaired quality of life (p < 0.001). Late seizures occurred in 43% of patients with postoperative seizures. The small sub-cohort of patients (2.7%) with a STAMPE2 score of more than six points had a significantly increased risk for seizures (p < 0.001, total risk 70%). We concluded that besides distinct risk factors, high scores of the modified STAMPE2 score could estimate the risk of postoperative seizures. However, it seems not transferable to our cohort