To evaluate the use of medical ultrasound combined with transfer learning to improve the diagnostic accuracy of pilomatricoma in children.

A dataset was constructed from ultrasound images obtained at the Affiliated Children's Hospital of Jiangnan University (Wuxi Children's Hospital) between December 2019 and February 2025. The dataset included 468 images of pilomatricoma, 66 images of epidermoid cysts, and 28 images of sebaceous adenomas. Four widely used deep learning models (AlexNet, ResNet18, ResNet50, and ResNet101) were pre-trained on large image datasets and subsequently fine-tuned via transfer learning for the specific task of pediatric pilomatricoma diagnosis. Diagnostic performance was assessed using four metrics: precision, recall, F1 score, and specificity. A comprehensive score (maximum 4 points), calculated as the sum of these four indicators, was used to evaluate the performance of each transfer learning model.

For the transfer learning models based on AlexNet, ResNet18, ResNet50, and ResNet101, the achieved results were as follows: precision: 97.4%, 94.9%, 98.6%, and 96.2%; recall: 100%, 98.7%, 96.0%, and 100%; F1 score: 98.7%, 96.7%, 97.3%, and 98.0%; specificity: 86.7%, 73.3%, 93.3%, and 80.0%, respectively. All four transfer learning models demonstrated superior diagnostic performance, significantly reducing the rates of misdiagnosis and missed diagnosis rates compared to conventional ultrasound interpretation. The comprehensive quantitative evaluation scores for the four models were 3.9, 3.7, 3.7, and 3.8, respectively, indicating strong generalization ability. The AlexNet-based model achieved the best overall performance.

Transfer learning-based deep learning models can provide more accurate ultrasound diagnostic information for pediatric pilomatricoma compared with current manual methods.