Preoperative differentiation of Wilms tumor and neuroblastoma on pediatric abdominal computed tomography (CT) images may be challenging because of overlapping imaging features. We aimed to develop an artificial intelligence-assisted lesion-localization model for exploratory diagnostic support in this differential setting. In this single-center, retrospective, image-level study, a YOLO26s detector was trained on preoperative contrast-enhanced CT PNG images with histopathology-anchored labels. The dataset comprised 3553 images, including 2103 lesion-positive images and 1450 background-negative images, partitioned into training, validation, and test subsets. On the held-out test set, the model achieved a precision of 0.954, a recall of 0.951, an mAP@0.5 of 0.977, and an mAP@0.5:0.95 of 0.732. Class-specific mAP@0.5:0.95 values were 0.734 for neuroblastoma and 0.730 for Wilms tumor. At the image level, tumor-present versus background-negative discrimination yielded 99.5% sensitivity, 89.0% specificity, a 93.0% positive predictive value, a 99.2% negative predictive value, and 95.3% accuracy. YOLO26s showed strong within-dataset performance for lesion localization and differential support between Wilms tumor and neuroblastoma.
Akay et al. (Mon,) studied this question.