Staged Unimodal-to-Multimodal AI Analysis of Histopathology, CT/MRI, and Multiplex Tissue Imaging for Perioperative Risk Prediction in Colorectal Cancer (KIA-Korekt)
Staged Unimodal-to-Multimodal AI Analysis of Histopathology, CT/MRI, and Multiplex Tissue Imaging for Perioperative Risk Prediction in Colorectal Cancer (KIA-Korekt)
Perioperative complications following surgery for colorectal cancer (CRC) represent a major cause of postoperative morbidity and mortality. Existing risk stratification tools lack the precision to capture the complex biological and morphological factors that determine individual patient vulnerability. Artificial intelligence (AI)-based analysis of medical imaging data offers a promising approach to improve preoperative risk prediction.
The KIA-Korekt study investigates whether perioperative complications in CRC patients can be predicted using multimodal AI-based image analysis. Three complementary imaging modalities are integrated: digital histopathology (haematoxylin-eosin whole-slide images, H&E-WSIs), preoperative CT and MRI radiomics, and multiplex tissue imaging (mTI) including multiplex immunohistochemistry (mIHC) and imaging mass cytometry (IMC).
The study includes a retrospective cohort of approximately 750 CRC patients treated between 2011 and 2021, and a prospective validation cohort of approximately 210 patients recruited from 2026 to 2028. Deep learning and radiomic feature extraction pipelines are applied to all modalities individually and in multimodal combination. Predicted outcomes include anastomotic leakage, wound infection, sepsis, ICU admission, and in-hospital mortality within 30 days of surgery.
The study is conducted at the University Hospital Brandenburg, Brandenburg Medical School Theodor Fontane, in collaboration with the Department of Computational Pathology, TU Dresden.
Colorectal cancer (CRC) is one of the most prevalent malignancies worldwide. Despite advances in surgical technique and perioperative care, short-term postoperative complications remain frequent and substantially impact patient quality of life, healthcare costs, and long-term prognosis. These complications include anastomotic leakage, wound infection, sepsis, thromboembolic events, and in-hospital mortality. Existing clinical risk scores (ASA, POSSUM) provide only limited individualised risk stratification and do not incorporate imaging-derived biological markers.
The KIA-Korekt study addresses this gap by developing and validating AI-based predictive models for perioperative complications in CRC, integrating three complementary imaging modalities:
Digital histopathology: Haematoxylin-eosin stained whole-slide images (H&E-WSIs) from surgical resection specimens and preoperative biopsies are analysed using attention-based multiple instance learning (MIL) and convolutional neural networks (CNNs), building on established pipelines from the Department of Computational Pathology, TU Dresden (AG Kather).
Radiology: Preoperative CT and MRI images are processed using automated segmentation (TotalSegmentator, nnU-Net) and radiomic feature extraction (PyRadiomics). Features are derived from the primary tumour, psoas muscle (sarcopenia), and visceral/subcutaneous fat compartments. A dedicated multi-metric quality control pipeline ensures stable imaging data representations across scanners and acquisition protocols.
Multiplex tissue imaging (mTI): Multiplex immunohistochemistry with multispectral imaging (mIHC-MSI) and imaging mass cytometry (IMC) are applied to formalin-fixed paraffin-embedded tumour tissue to characterise immune and stromal cell populations, marker expression intensities, and spatial distribution patterns within the tumour microenvironment.
Unimodal models are developed and validated separately for each modality. Multimodal integration is performed using feature-level fusion, late fusion, and multimodal multiple-instance learning with cross-attention mechanisms. Model performance is evaluated using AUC-ROC, calibration plots, Brier scores, and Decision Curve Analysis. Interpretability is assessed using SHAP values and attention heatmaps.
The study employs a mixed retrospective (n=750, 2011-2021) and prospective validation (n=210, 2026-2028) cohort design. The retrospective cohort provides the basis for model development and internal cross-validation; the prospective cohort enables real-world external validation under clinical conditions.
A comprehensive patient-level macro-micro correlation analysis investigates associations between radiological imaging phenotypes and microscopic histopathological and immunological characteristics derived from the same tumours, enabling unique integrative biological insights.
The study is funded by the European Union and the State of Brandenburg (HealthTranslateBB/ERDF) and the German Research Foundation (DFG). Ethics approval has been granted by the ethics committee of the Brandenburg Medical School Theodor Fontane. All prospective participants provide written informed consent. Retrospective data are processed in pseudonymised form in accordance with GDPR.
Results will be disseminated through peer-reviewed open-access publications and national and international conference presentations.
Inclusion Criteria:
Exclusion Criteria:
melissa.schadl@mhb-fontane.de+4915207809673