Arjen Cleven

UMC Groningen

In short, Arjen Cleven's research entails the following:

Classification of craniofacial bone tumors using deep learning algorithms

Fibro-osseous bone tumors occurring in the craniofacial skeleton represent a group of tumors with diverse clinical behavior. The pathogenesis of most of these tumors is not completely understood. Although their histological appearance may be similar, the clinical behavior of these entities varies from indolent to aggressive growth and/or metastatic spread. Therefore, an accurate diagnosis is crucial for adequate treatment and prognosis.

Daily practice shows that radiological as well as histomorphological features largely overlap between these different entities. Therefore, correlation of clinical features, radiology and histomorphology are extremely important to reliably diagnose fibro-osseous bone tumors. But even after evaluation by a multidisciplinary team of experts of all the available information including genetics, the final diagnosis remains often uncertain in a significant number of cases, especially on limited biopsy material.

Recognizing these limitations, we propose an innovative extra approach to address the diagnostic challenges associated with fibro-osseous tumors in craniofacial bones. Our study pioneers an advanced deep learning model for tumor classification using whole slide images (WSIs) derived from haematoxylin and eosin-stained tissue samples. The model utilizes state-of-the-art foundational models to extract morphological and textural features from image patches. These features are then aggregated using an attention mechanism, which assigns relevance scores to each patch, enabling precise tumor classification. This approach enhances interpretability by not only providing accurate results but also highlighting the image regions that were most influential in the decision-making process.

Our study is an international and multicentered collaboration consisting of a well-characterized discovery cohort of craniofacial bone tumors with available molecular data and WSIs and two external validation cohorts including the Doesak registry from the University Hospital Basel Switzerland and craniofacial bone tumors from the Dutch Bone Tumor Commission.

We believe that using deep learning on WSIs in an integrated multidisciplinary diagnostic approach will increase diagnostic accuracy for fibro-osseous tumors of craniofacial bone,especially for osteosarcomas.

Short summary of progress/results January 2026

In the first year of the project, we successfully added 186 additional craniofacial bone tumor cases to the original discovery cohort, creating a dataset of more than 300 craniofacial bone tumor cases for reprocessing in our classification model. We used a new universal whole slide image (WSI) patching library developed in our research facility; several types of stain normalization (e.g. Macenko) and an AI tissue segmentation model (MobileNetV3) were implemented. As an extension of the earlier preliminary work (which was based on convolutional neural networks), we developed a new Multiple Instance Learning (MIL) method using the foundation model UNI-2 [1]. Currently a manuscript is being prepared to publish our first results which is in line with our objective of the specified time.

Furthermore, to strengthen our model we are planning to incorporate CT scans into our multimodal AI framework and in addition in collaboration with Tumor Genomic Analysis Core group we are evaluating the possibility to utilize existing molecular data. To visualize which histopathological features are important for model prediction, we aim to create an explainability framework to publish with our model.