AB025. Deep learning model predicts recurrence in thymic epithelial tumours: a multicentre study
8 affiliations
- 1Department of Pathology and Clinical Bioinformatics, Erasmus Medical Centre, Rotterdam, The Netherlands
- 2Department of Radiation Oncology (Maastro), GROW Research Institute for Oncology and Reproduction, Maastricht University Medical Centre+, Maastricht, The Netherlands
- 3Department of Respiratory Medicine, Erasmus Medical Centre, Rotterdam, The Netherlands
- 4Department of Pulmonology, Maastricht UMC+, Maastricht, the Netherlands
- 5Department of Pathology, Maastricht UMC+, Maastricht, the Netherlands
- 6Department of Radiotherapy, Erasmus MC Cancer Institute, Rotterdam, The Netherlands
- 7Department of Pathology, Centre Hospitalier Universitaire de Lyon, Lyon, France
- 8Department of Medical Oncology, Curie Institute, Paris, France
Abstract
Background: Thymic epithelial tumours (TETs) are the most common primary neoplasms of the anterior mediastinum. Despite complete surgical resection (R0), a significant proportion of patients experience recurrence, highlighting the need for improved prognostic stratification. Morphological patterns recognised in clinical practice may have prognostic value, and this study applies deep learning (DL) to haematoxylin and eosin (H&E) whole slide images (WSIs) to predict recurrence in TETs, assisting pathologists in risk assessment and enabling precise treatment planning.
Methods: A multicentre cohort of resected TET patients with complete (R0) resection and no distant metastasis was assembled from Erasmus Medical Centre (EMC), Maastricht University Medical Centre (MUMC+) and the TCGA public dataset, with an external validation cohort from Hospices Civils de Lyon (HCL). WSIs were annotated in QuPath and sampled at 20× magnification into 512×512 patches. HCL WSIs, stained with haematoxylin-eosin saffron, underwent Vahadane normalisation to harmonise staining across centres. Patch embeddings from a pretrained Phikon encoder were aggregated through a multi-instance learning for patient level recurrence prediction, using stratified three-fold cross-validation. GradCAM applied to the most contributing patches generated heatmaps of the areas of highest model activation.
Results: Recurrence occurred in 19 of 298 patients (6.4%) internally and 5 of 95 (5.3%) externally. The model achieved an Area Under the Curve (AUC) of 0.85, a balanced accuracy (BAcc) of 0.71, and sensitivity of 79% internally. Performance was preserved externally (AUC 0.70, BAcc 0.62, sensitivity 80%) despite domain shift. In true positive cases, heatmaps consistently localised to tumour tissue adjacent to fibrous septa, capsular structures and vascular interfaces, with additional focal activation in areas of increased cellular heterogeneity in tumours lacking prominent separation.
Conclusions: The DL model can identify TET patients at risk of recurrence following R0 resection and demonstrates generalisability across institutions with differing tissue preparation protocols. Heatmaps corresponded to septal, capsular and perivascular structures, sites classically associated with tumour invasion and recurrence risk in TETs, suggesting the model may capture histological correlates of invasive behaviour. Future work will focus on larger datasets, multimodal architectures integrating clinical and imaging data, and further external validation.
Related
- AB024. Mixed histology is a prognostic factor for recurrence in completely resected thymic epithelial tumours: a multicentre retrospective study
- An interpretable machine-learning signature predicts early triple-negative breast cancer recurrence but adds no prognostic value beyond clinicopathological factors in external validation
- [177Lu]Lu-edotreotide versus everolimus for gastroenteropancreatic neuroendocrine tumours (COMPETE): a phase 3, multicentre, randomised, open-label, superiority trial
- Routine MDM2 FISH Testing and Preoperative Biopsy Reduces Recurrence Rate of Atypical Lipomatous Tumours
- AB027. Paraneoplastic diseases in patients with thymic epithelial tumours: findings from the nation-wide Danish thymic epithelial tumor database.
BibTeX
@article{dominguez2026ab025,
title = {{AB025. Deep learning model predicts recurrence in thymic epithelial tumours: a multicentre study}},
author = {Anna Salut Esteve Domínguez and Stéphanie Peeters and Marlou Dimmers and Daphne Dumoulin and Florit Marcuse and Myrurgia Abdul-Hamid and Rianne Fijten and Yunlei Li and Andrew Stubbs and Dirk De Ruysscher and Lara Chalabreysse and Nicolas Girard and Jan von der Thüsen and Farhan Akram},
journal = {Mediastinum},
year = {2026},
month = oct,
doi = {10.21037/med-26-ab025},
url = {https://doi.org/10.21037/med-26-ab025},
note = {CC BY-NC-ND 4.0},
}
RIS
TY - JOUR
TI - AB025. Deep learning model predicts recurrence in thymic epithelial tumours: a multicentre study
AU - Anna Salut Esteve Domínguez
AU - Stéphanie Peeters
AU - Marlou Dimmers
AU - Daphne Dumoulin
AU - Florit Marcuse
AU - Myrurgia Abdul-Hamid
AU - Rianne Fijten
AU - Yunlei Li
AU - Andrew Stubbs
AU - Dirk De Ruysscher
AU - Lara Chalabreysse
AU - Nicolas Girard
AU - Jan von der Thüsen
AU - Farhan Akram
PY - 2026
DA - 2026/10/30
JO - Mediastinum
DO - 10.21037/med-26-ab025
UR - https://doi.org/10.21037/med-26-ab025
AB - Background: Thymic epithelial tumours (TETs) are the most common primary neoplasms of the anterior mediastinum. Despite complete surgical resection (R0), a significant proportion of patients experience recurrence, highlighting the need for improved prognostic stratification. Morphological patterns recognised in clinical practice may have prognostic value, and this study applies deep learning (DL) to haematoxylin and eosin (H&E) whole slide images (WSIs) to predict recurrence in TETs, assisting pathologists in risk assessment and enabling precise treatment planning. Methods: A multicentre cohort of resected TET patients with complete (R0) resection and no distant metastasis was assembled from Erasmus Medical Centre (EMC), Maastricht University Medical Centre (MUMC+) and the TCGA public dataset, with an external validation cohort from Hospices Civils de Lyon (HCL). WSIs were annotated in QuPath and sampled at 20× magnification into 512×512 patches. HCL WSIs, stained with haematoxylin-eosin saffron, underwent Vahadane normalisation to harmonise staining across centres. Patch embeddings from a pretrained Phikon encoder were aggregated through a multi-instance learning for patient level recurrence prediction, using stratified three-fold cross-validation. GradCAM applied to the most contributing patches generated heatmaps of the areas of highest model activation. Results: Recurrence occurred in 19 of 298 patients (6.4%) internally and 5 of 95 (5.3%) externally. The model achieved an Area Under the Curve (AUC) of 0.85, a balanced accuracy (BAcc) of 0.71, and sensitivity of 79% internally. Performance was preserved externally (AUC 0.70, BAcc 0.62, sensitivity 80%) despite domain shift. In true positive cases, heatmaps consistently localised to tumour tissue adjacent to fibrous septa, capsular structures and vascular interfaces, with additional focal activation in areas of increased cellular heterogeneity in tumours lacking prominent separation. Conclusions: The DL model can identify TET patients at risk of recurrence following R0 resection and demonstrates generalisability across institutions with differing tissue preparation protocols. Heatmaps corresponded to septal, capsular and perivascular structures, sites classically associated with tumour invasion and recurrence risk in TETs, suggesting the model may capture histological correlates of invasive behaviour. Future work will focus on larger datasets, multimodal architectures integrating clinical and imaging data, and further external validation.
SN - 2522-6711
ER -