Contemporary deep learning-based decision systems are well-known for requiring high-volume datasets in order to produce generalized, reliable, and high-performing models. However, the collection of such datasets is challenging, requiring time-consuming processes involving also expert clinicians with limited time. In addition, data collection often raises ethical and legal issues and depends on costly and invasive procedures. Deep generative models such as generative adversarial networks and variational autoencoders can capture the underlying distribution of the examined data, allowing them to create new and unique instances of samples. This study aims to shed light on generative data augmentation techniques and corresponding best practices. Through in-depth investigation, we underline the limitations and potential methodology pitfalls from critical standpoint and aim to promote open science research by identifying publicly available open-source repositories and datasets.

Prostate cancer (PCa) is currently the second most prevalent cancer among men. Accurate diagnosis of PCa can provide effective treatment for patients and reduce mortality. Previous works have merely focused on either lesion detection or lesion classification of PCa from magnetic resonance imaging (MRI). In this work we focus on a critical, yet underexplored task of the PCa clinical workflow: distinguishing cases with cancer presence (pathologically confirmed PCa patients) from conditions with no suspicious PCa findings (no cancer presence). To this end, we conduct large-scale experiments for this task for the first time by adopting and processing the multi-centric ProstateNET Imaging Archive which contains more than 6 million image representations of PCa from more than 11,000 PCa cases, representing the largest collection of PCa MR images. Bi-parametric MR (bpMRI) images of 4504 patients alongside their clinical variables are used for training, while the architectures are evaluated on two hold-out test sets of 975 retrospective and 435 prospective patients. Our proposed multi-encoder-cross-attention-fusion architecture achieved a promising area under the receiver operating characteristic curve (AUC) of 0.91. This demonstrates our method’s capability of fusing complex bi-parametric imaging modalities and enhancing model robustness, paving the way towards the clinical adoption of deep learning models for accurately determining the presence of PCa across patient populations.

Accurate patient-level risk stratification of clinically significant prostate cancer (csPCa) on biparametric MRI remains challenging, as most existing approaches require costly lesion-level annotations or have not been evaluated in large temporally separated cohorts. We developed and validated a multimodal deep learning framework combining T2-weighted (T2w), diffusion-weighted (DWI), and apparent diffusion coefficient (ADC) volumes with age, prostate-specific antigen (PSA), and PSA density (PSAD) for patient-level csPCa classification, trained without lesion-level annotations. The study included a retrospective development cohort of 3, 939 examinations and an independent temporal validation cohort comprising 1, 409 prospectively acquired examinations from 13 centers. The model employs three sequence-specific encoders with asymmetric cross-attention fusion and a two-stage transfer-learning strategy. Adding demographic and clinical variables consistently improved performance over MRI-only models, with PSAD emerging as the most informative complementary factor, primarily by improving specificity. The best-performing model—combining MRI, age, PSA, and PSAD with cross-attention—achieved a mean AUC of $$0.765\!\pm \!0.006$$ on the retrospective held-out test set and $$0.741\!\pm \!0.006$$ on the independent temporal validation cohort. Post hoc subgroup analyses suggested broadly stable rank-order discrimination (AUC) across most strata, though fixed-threshold sensitivity and specificity varied substantially, while Grad-CAM maps offered a qualitative indication of more frequent overlap with suspicious regions under cross-attention. These findings support PSAD-informed multimodal patient-level models as a viable tool for refined csPCa risk stratification on biparametric MRI, with potential to reduce false-positive classification of ISUP grade group 1 disease as clinically significant.