AI Analysis of CT Scans to Classify Ovarian Tumours as Benign, Borderline or Malignant Before Surgery

* Background and rationale Preoperative distinction between benign, borderline and malignant ovarian tumours determines referral, surgical approach and extent of surgery. Ultrasound-based models such as IOTA-ADNEX and the Risk of Malignancy Index (RMI) are the current triage tools, but their performance varies between populations and users. CT is often performed in indeterminate cases or for staging. There are no validated diagnostic criteria for CT, and interpretation depends on the reader. Published CT-based computer-aided diagnosis (CADx) models are mostly single-centre, binary (benign versus malignant), dependent on manual tumour segmentation and rarely externally validated. Borderline tumours need different management but are usually excluded or merged with malignant tumours. Earlier work from the Catharina Cancer Institute and Eindhoven University of Technology (TU/e) showed considerable domain shift between hospitals in a multicentre CT dataset (Koch et al., 2026). * Index test The index test is a CADx pipeline built on SPECTRE, a CT foundation model developed at TU/e and pre-trained with self-supervised learning on more than 250,000 CT scans. SPECTRE is fine-tuned with parameter-efficient adaptation. The pre-trained encoder is kept frozen or partially unfrozen, and lightweight task heads are trained for tumour localisation and three-class classification. The pipeline works directly on the CT volume. Tumour localisation is learned within the model, so no tumour masks or bounding boxes are needed for training or evaluation. An uncertainty quantification module produces calibrated confidence estimates alongside each prediction. The overall architecture is fixed. Training details such as the fine-tuning schedule, loss weighting and calibration method may be adapted based on interim results. Radiomics-based machine learning (handcrafted features with classifiers such as logistic regression, support vector machines, random forests or gradient boosting) is used only as a benchmark comparator. The model is applied retrospectively and its output is never available to treating clinicians. * Reference standard Definitive postoperative histopathology, classified as benign, borderline or malignant according to the 2020 WHO Classification of Female Genital Tumours. Ambiguous cases are adjudicated by an experienced gynaecological pathologist to keep labelling consistent across centres. In bilateral tumours with discordant histology, a predefined rule is applied (for example, selection of the lesion with the highest malignant potential). * Data acquisition and preprocessing Participating centres identify eligible patients and pseudonymise and code imaging, clinical and pathology data locally before transfer under formal data-sharing agreements. The key linking codes to identities stays at the local centre. Images are exported in DICOM format and may be converted to volumetric formats (NIfTI or NRRD). CT preprocessing includes intensity windowing (for example -100 to +300 HU), normalisation and voxel resampling to a standard resolution. Scanner manufacturer and model, slice thickness, contrast phase and reconstruction kernel are recorded for analysis of domain shift. For the MRI pilot, T2-weighted, T1-weighted with and without fat suppression, diffusion-weighted with ADC maps and contrast-enhanced T1-weighted sequences are used when available. * Model development and validation Within the development cohort, stratified cross-validation (or an equivalent resampling method adapted to the final sample size) is used, stratified by outcome class. Hyperparameter tuning and model selection take place only within the training data of each resampling iteration. All scans and lesions from one patient are kept in the same split to prevent data leakage. External validation uses one or more centres that are excluded entirely from model development. The allocation of centres to development and external validation is documented before any performance analysis. Calibration methods such as Platt scaling or temperature scaling may be applied. * Statistical analysis Discrimination is assessed with one-vs-rest AUROC per class and macro- and micro-averaged multiclass AUROC. Classification metrics include sensitivity, specificity, positive and negative predictive values, F1-score, balanced accuracy and confusion matrices. 95% confidence intervals are estimated by bootstrapping where appropriate. Calibration is assessed with calibration curves and the Brier score. Net benefit across clinically relevant risk thresholds is assessed with decision curve analysis. Comparison with RMI, IOTA-ADNEX, radiologist reports and, in the MRI pilot, O-RADS MRI is restricted to patients in whom the comparator is available or can be reconstructed. Paired comparisons use the DeLong test for ROC curves, McNemar's test for paired proportions, kappa statistics for agreement and the Brier score for probabilistic predictions. Baseline characteristics are compared with chi-squared or Fisher's exact tests and t-tests or Mann-Whitney U tests as appropriate. Two-sided p-values are reported with a nominal significance level of 0.05, with emphasis on effect sizes and confidence intervals. Missing clinical data are described and not imputed. Model behaviour is explored with attention and saliency maps (for example Grad-CAM). * Exploratory analyses Where data can be linked in coded form and are covered by ethics approval and data-sharing agreements, exploratory analyses will assess whether adding blood-based biomarkers (CA-125, CA-125 glycovariants, CEA, ctDNA, proteomics) or radiologist report content to CT imaging changes classification performance. In a subset with digitised histopathology slides, AI-based analysis of whole slide images will be explored to find histological correlates of CT-based tumour characteristics. These analyses do not change the primary imaging-based objective. * Reporting Results will be reported following TRIPOD+AI, supplemented by CLAIM and STROBE where applicable.