We built a general-purpose radiological visual representation using the large-scale adult CT dataset CT-3M, and introduced a Gram-anchoring mechanism that employs a frozen adult pretrained model as a structural teacher to guide domain alignment of local topological structures on unlabeled pediatric CT data. Combined with a multi-scale feature aggregation strategy and a lightweight Primus decoder, downstream segmentation tasks were evaluated on a public pediatric CT dataset. Based on case-wise paired results, we compared the mean Dice similarity coefficient (DSC) and mean intersection over union (IoU) between our model and the baseline nnU-Net using the Wilcoxon signed-rank test, and computed the relative performance improvements.

A total of 867 abdominal CT imaging cases were collected, constituting a pretraining dataset comprising 367 588 two-dimensional CT slices. On the public Pediatric-CT-SEG dataset (359 cases), our model achieved a mean DSC of (71.38±1.08)% and a mean IoU of (63.73±1.01)%, representing improvements of 3.22% and 3.59% over the baseline nnU-Net, respectively, with statistically significant differences (P < 0.05). Stable improvements in mean DSC were also observed for small-volume or boundary-ambiguous organs, including the duodenum (5.78%), pancreas (4.69%), left adrenal gland (2.89%), right adrenal gland (1.12%), and gallbladder (1.86%). Ablation experiments demonstrated that DSC improved by 1.44%, 3.66%, 5.78%, and 6.45% following adult pretraining, pediatric domain adaptation, high-resolution adaptation, and multi-scale feature aggregation, respectively.

The self-supervised pretraining framework proposed in this study effectively alleviates the domain shift between adult and pediatric abdominal CT images, significantly enhances segmentation accuracy for pediatric abdominal multi-organs-particularly small organs and structures with complex boundaries-and provides a reliable technical solution for intelligent pediatric imaging analysis in scenarios with limited annotated data.