Maarten de Rooij

Bram de Wilde, Anindo Saha, Maarten de Rooij et al.

Diffusion models for text-to-image generation, known for their efficiency, accessibility, and quality, have gained popularity. While inference with these systems on consumer-grade GPUs is increasingly feasible, training from scratch requires large captioned datasets and significant computational resources. In medical image generation, the limited availability of large, publicly accessible datasets with text reports poses challenges due to legal and ethical concerns. This work shows that adapting pre-trained Stable Diffusion models to medical imaging modalities is achievable by training text embeddings using Textual Inversion. In this study, we experimented with small medical datasets (100 samples each from three modalities) and trained within hours to generate diagnostically accurate images, as judged by an expert radiologist. Experiments with Textual Inversion training and inference parameters reveal the necessity of larger embeddings and more examples in the medical domain. Classification experiments show an increase in diagnostic accuracy (AUC) for detecting prostate cancer on MRI, from 0.78 to 0.80. Further experiments demonstrate embedding flexibility through disease interpolation, combining pathologies, and inpainting for precise disease appearance control. The trained embeddings are compact (less than 1 MB), enabling easy data sharing with reduced privacy concerns.

Michelle Stegeman, Lena Philipp, Fennie van der Graaf et al.

Medical foundation models show promise to learn broadly generalizable features from large, diverse datasets. This could be the base for reliable cross-modality generalization and rapid adaptation to new, task-specific goals, with only a few task-specific examples. Yet, evidence for this is limited by the lack of public, standardized, and reproducible evaluation frameworks, as existing public benchmarks are often fragmented across task-, organ-, or modality-specific settings, limiting assessment of cross-task generalization. We introduce UNICORN, a public benchmark designed to systematically evaluate medical foundation models under a unified protocol. To isolate representation quality, we built the benchmark on a novel two-step framework that decouples model inference from task-specific evaluation based on standardized few-shot adaptation. As a central design choice, we constructed indirectly accessible sequestered test sets derived from clinically relevant cohorts, along with standardized evaluation code and a submission interface on an open benchmarking platform. Performance is aggregated into a single UNICORN Score, a new metric that we introduce to support direct comparison of foundation models across diverse medical domains, modalities, and task types. The UNICORN test dataset includes data from more than 2,400 patients, including over 3,700 vision cases and over 2,400 clinical reports collected from 17 institutions across eight countries. The benchmark spans eight anatomical regions and four imaging modalities. Both task-specific and aggregated leaderboards enable accessible, standardized, and reproducible evaluation. By standardizing multi-task, multi-modality assessment, UNICORN establishes a foundation for reproducible benchmarking of medical foundation models. Data, baseline methods, and the evaluation platform are publicly available via unicorn.grand-challenge.org.

Joeran S. Bosma, Anindo Saha, Matin Hosseinzadeh et al.

Deep learning-based diagnostic performance increases with more annotated data, but large-scale manual annotations are expensive and labour-intensive. Experts evaluate diagnostic images during clinical routine, and write their findings in reports. Leveraging unlabelled exams paired with clinical reports could overcome the manual labelling bottleneck. We hypothesise that detection models can be trained semi-supervised with automatic annotations generated using model predictions, guided by sparse information from clinical reports. To demonstrate efficacy, we train clinically significant prostate cancer (csPCa) segmentation models, where automatic annotations are guided by the number of clinically significant findings in the radiology reports. We included 7,756 prostate MRI examinations, of which 3,050 were manually annotated. We evaluated prostate cancer detection performance on 300 exams from an external centre with histopathology-confirmed ground truth. Semi-supervised training improved patient-based diagnostic area under the receiver operating characteristic curve from $87.2 \pm 0.8\%$ to $89.4 \pm 1.0\%$ ($P<10^{-4}$) and improved lesion-based sensitivity at one false positive per case from $76.4 \pm 3.8\%$ to $83.6 \pm 2.3\%$ ($P<10^{-4}$). Semi-supervised training was 14$\times$ more annotation-efficient for case-based performance and 6$\times$ more annotation-efficient for lesion-based performance. This improved performance demonstrates the feasibility of our training procedure. Source code is publicly available at github.com/DIAGNijmegen/Report-Guided-Annotation. Best csPCa detection algorithm is available at grand-challenge.org/algorithms/bpmri-cspca-detection-report-guided-annotations/.

Alessa Hering, Sarah de Boer, Anindo Saha et al.

The PI-CAI (Prostate Imaging: Cancer AI) challenge led to expert-level diagnostic algorithms for clinically significant prostate cancer detection. The algorithms receive biparametric MRI scans as input, which consist of T2-weighted and diffusion-weighted scans. These scans can be misaligned due to multiple factors in the scanning process. Image registration can alleviate this issue by predicting the deformation between the sequences. We investigate the effect of image registration on the diagnostic performance of AI-based prostate cancer diagnosis. First, the image registration algorithm, developed in MeVisLab, is analyzed using a dataset with paired lesion annotations. Second, the effect on diagnosis is evaluated by comparing case-level cancer diagnosis performance between using the original dataset, rigidly aligned diffusion-weighted scans, or deformably aligned diffusion-weighted scans. Rigid registration showed no improvement. Deformable registration demonstrated a substantial improvement in lesion overlap (+10% median Dice score) and a positive yet non-significant improvement in diagnostic performance (+0.3% AUROC, p=0.18). Our investigation shows that a substantial improvement in lesion alignment does not directly lead to a significant improvement in diagnostic performance. Qualitative analysis indicated that jointly developing image registration methods and diagnostic AI algorithms could enhance diagnostic accuracy and patient outcomes.

Anindo Saha, Joeran S. Bosma, Jasper J. Twilt et al.

In this intercontinental, confirmatory study, we include a retrospective cohort of 22,481 MRI examinations (21,288 patients; 46 cities in 22 countries) to train and externally validate the PI-CAI-2B model, i.e., an efficient, next-generation iteration of the state-of-the-art AI system that was developed for detecting Gleason grade group $\geq$2 prostate cancer on MRI during the PI-CAI study. Of these examinations, 20,471 cases (19,278 patients; 26 cities in 14 countries) from two EU Horizon projects (ProCAncer-I, COMFORT) and 12 independent centers based in Europe, North America, Asia and Africa, are used for training and internal testing. Additionally, 2010 cases (2010 patients; 20 external cities in 12 countries) from population-based screening (STHLM3-MRI, IP1-PROSTAGRAM trials) and primary diagnostic settings (PRIME trial) based in Europe, North and South Americas, Asia and Australia, are used for external testing. Primary endpoint is the proportion of AI-based assessments in agreement with the standard of care diagnoses (i.e., clinical assessments made by expert uropathologists on histopathology, if available, or at least two expert urogenital radiologists in consensus; with access to patient history and peer consultation) in the detection of Gleason grade group $\geq$2 prostate cancer within the external testing cohorts. Our statistical analysis plan is prespecified with a hypothesis of diagnostic interchangeability to the standard of care at the PI-RADS $\geq$3 (primary diagnosis) or $\geq$4 (screening) cut-off, considering an absolute margin of 0.05 and reader estimates derived from the PI-CAI observer study (62 radiologists reading 400 cases). Secondary measures comprise the area under the receiver operating characteristic curve (AUROC) of the AI system stratified by imaging quality, patient age and patient ethnicity to identify underlying biases (if any).