Three-stage computer-vision pipeline for ovarian carcinoma diagnosis from transvaginal ultrasound in postmenopausal women: dataset, annotations, features, and model weights

This Zenodo record contains the de-identified data, annotations, derived features, and trained model weights that accompany the manuscript "Three-stage computer-vision pipeline for ovarian carcinoma diagnosis from transvaginal ultrasound in postmenopausal women." ETHICS AND DE-IDENTIFICATION This retrospective study was approved by the Meir Medical Center Helsinki Committee (approval number 0005-23-MMC). All images and metadata are fully de-identified: patient identifiers were removed, no device overlays or on-screen annotations are present in the released images, and patients were assigned anonymised study codes that bear no relation to medical record numbers. Calendar dates were stripped from all filenames and CSV columns; files are organised by patient code and 1-based visit ordinal (for example, 141.v2.3.png), conveying only the relative ordering of visits and no absolute timing information. The Helsinki Committee approval explicitly covers public sharing of the de-identified dataset. DATASET OVERVIEW - Patients: 260 postmenopausal women evaluated at Meir Medical Center (2012 to 2023).- Diagnostic groups: sonographically normal 131 patients / 529 images; benign 68 / 291; malignant 61 / 262.- Total images: 1,082 transvaginal 2D B-mode ultrasound images, standardised to 1,280 x 1,280 pixels (PNG).- Patient-level split: 70% training, 15% validation, 15% test, stratified by patient and class. Stage I and Stage III were split independently, each over its own denominator, so their test cohorts are different patients.- Ground truth: histopathological confirmation for benign and malignant cases; clinical assessment for normal controls. CONTENTS - images_normal.zip / images_benign.zip / images_malignant.zip: 1,082 padded PNG files organised by diagnostic class.- yolo_labels.zip: cyst bounding-box annotations in YOLO format, with the benign and malignant labels as the two classes (553 .txt files for the pathological cohort).- clinical.zip: XLSX files with 15 clinical features per patient (5 continuous, 10 binary indicators) for the pathological cohort (n = 129). CA-125 was measured preoperatively in 124 of the 129; in the remaining five, all benign, the value is stored as 0, which is not physiologically possible, and the manuscript treats those as missing.- morphological_features.zip: 11 features per image (8 geometric, 3 intensity) computed from Segment Anything Model (SAM ViT-H) cyst masks. Two sets are released: masks prompted with the manual annotation box, and masks prompted with the Stage II detector box. The second set is the input of Configuration 6; the two are correlated but not interchangeable.- detector_boxes.zip: the Stage II bounding box of each validation and test image, with the full-frame fallback where the detector found nothing.- splits.zip: the per-patient partition summary and the per-image Stage I and Stage III partitions, in evaluation order.- predictions.zip: per-image and per-patient predictions of Stage I and of both reported Stage III configurations.- stage1_run_files.zip: configuration and metrics of the reported Stage I run.- stage1_efficientnet_b7.pth: EfficientNet-B7 weights for Stage I (normal versus pathological triage, trained with a clinical CAM-loss).- stage2_yolov8m.pt: YOLOv8m weights for Stage II (cyst localisation).- stage3_config2_baseline.pth: Stage III Configuration 2 (ROI + CNN, image-only baseline).- stage3_config6_morphological_fusion.pth: Stage III Configuration 6 (ROI + CNN + 11 morphological features).- stage3_config6_feature_normalization.json: descriptor means and standard deviations, fitted on the training images only, required by Configuration 6.- source_code.zip: the companion code repository at commit 9e140b1, a documented release of the code used in the study (MIT License). REPORTED PERFORMANCE - Stage I (triage, held-out test set): image-level accuracy 0.973, patient-level accuracy 0.974; AUC 0.996 and 0.997. Stage I separates routine scans of sonographically normal ovaries from scans of evident adnexal masses in this single-centre cohort, and should not be read as a screening result.- Stage II (detection, validation set): mAP@0.5 = 0.746.- Stage III Configuration 2 (test set): patient-level accuracy 0.905, AUC 0.945 (95% CI 0.824 to 1.000), Cohen's kappa 0.809.- Stage III Configuration 6 (test set): patient-level accuracy 0.952, AUC 0.964 (95% CI 0.873 to 1.000), Cohen's kappa 0.905, with no malignant case missed.- Three gynaecologists scoring the same 21 test patients unaided reached 0.714, 0.571 and 0.524, and a majority vote across them 0.667. The test cohort is small: one patient moves patient-level accuracy by 4.8 percentage points, and most of the differences above are not statistically significant. The manuscript reports every confidence interval and paired test. SAM MASKS Per-image SAM masks are not redistributed in this record. The morphological features are derived deterministically by loading the bounding box, prompting Segment Anything (SAM ViT-H, public checkpoint sam_vit_h_4b8939.pth from Meta AI) with that box, and computing the features from the resulting binary mask. The full pipeline that regenerates the masks is provided in the companion source-code repository. INTENDED USE AND LIMITATIONS This dataset is released for research purposes only and must not be used to make clinical decisions. The models have not been validated externally at the time of release; external validation across centres, scanners, and patient populations is required before any clinical translation. Geometric features are reported in pixel units (no pixel-to-millimetre calibration was available across the multi-vendor cohort) and are therefore relative within this cohort. COMPANION SOURCE CODE Source code for data preprocessing, model training, inference, statistical analysis, and figure generation is provided at https://github.com/Amit-Brilant/ovarian-cancer-tvus-pipeline under the MIT License, and a snapshot of that repository is archived in this record as source_code.zip. LICENSE Data, annotations, features, predictions and model weights are released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license; the source code in source_code.zip under the MIT License. CITATION Please cite both this Zenodo record and the accompanying manuscript when using these data or models. Citation strings (BibTeX) are provided in the README.md file included in this record.