康峻宏黎阮國慶Le Nguyen Binh2026-05-062026-05-062025-07-05https://203.71.86.71/handle/123456789/9161學位別：博士 口試委員：康峻宏; 黎阮國慶; 蘇家玉; 王禎麒; 吳孟晃 關鍵字：deep learning、pediatric distal forearm fracture、AO/OTA classification、YOLO、pediatric elbow fracture、imaging detection、systematic review、meta-analysisBackground: Pediatric upper extremity fractures are the most common trauma in children, accounting for 60-80% of all fractures in children aged 3 months to 15 years. Distal forearm and elbow fractures are the most frequently encountered types. These fractures can lead to severe complications and significantly impact a child’s functional development in the long term. Therefore, early and accurate diagnosis and appropriate intervention are crucial in improving patient treatment outcomes. However, clinical diagnosis remains challenging due to the limited number of experienced physicians in pediatric fracture imaging. As a result, developing AI models based on deep learning could provide valuable assistance to clinicians in fracture diagnosis. Objectives: This study has two aims: 1) To develop a CNN model to detect and classify pediatric distal forearm fractures based on the AO/OTA classification system guidelines for pediatric fractures; 2) To systematically evaluate the performance of DL models in detecting pediatric elbow fractures. Methods: In the first part, we obtained images of wrist X-rays from the GRAZPEDWRI-DX dataset published by the Department for Pediatric Surgery of the University Hospital Graz between 2008 and 2018. The photos were labeled into 4 classes (FRM, FUM, FRE, FUE) in fracture position divided into metaphysis and epiphysis in both radius and ulna depending on AO/ATO classification in pediatric fracture. We trained a CNN object detection model based on YOLOv4 with a training set of 7006 images (n= 1809), 80% for training and 20% for validation. A randomly selected test set with 88 images (n=34) was used for evaluation the model and comparison to 2 readers , one orthopedist and one radiologist. Next in the second part, we made a systematic review and meta-analysis by searching PubMed (Medline), EMBASE, and IEEExplore for studies published until October 20, 2023. Studies utilizing a deep learning model for detecting elbow fractures in patients aging from 0 to 16 years old were included. Sensitivity, specificity, and AUC were obtained. This study was registered with PROSPERO with ID being CRD42023470558. Results: The overall mean average precision (mAP) on the validation set in 4 classes were 0.97, 0.92, 0.95 and 0.94, respectively. On the test set, the sensitivity were 0.864, 0.714, 0.882, and 0.894, respectively; the specificity was 0.88, 0.945, 0.971, and 0.985; the AUC was 0.87, 0.83, 0.93, 0.94, respectively. The mean reading time was 0.056 ± 0.025 with model performance. The best performance among three readers belonged to radiologist with mean AUC (0.922), followed by our model (0.892) and the orthopedist (0.83), P<0.05. In the second part, our search identified 22 studies, of which six studies were included in the meta-analysis. The pooled estimate of the sensitivity of DL models in pediatric elbow fracture detection was 0.93 (95% CI 0.91 - 0.96). Specificity values for fracture detection ranged from 0.84 to 0.92 across studies, with a pooled estimate of 0.89 (95% CI 0.85 - 0.92). The AUC of DL model in fracture detection ranged from 0.91 to 0.99, with a pooled estimate AUC of 0.95 (CI 95% 0.93 - 0.97). Further analysis revealed that the use of preprocessing methods and architecture of model's backbone. Conclusion: Our multiclass fracture detection model, based on the AO/OTA classification system, demonstrated high efficacy in identifying and classifying pediatric distal forearm fractures, providing valuable guidance for emergency clinicians. This can significantly reduce the risk of misdiagnosis and help non-specialized clinicians establish preliminary treatment protocols for pediatric cases. Additionally, deep learning models have shown outstanding accuracy in detecting pediatric elbow fractures. To achieve optimal efficiency, we recommend utilizing backbone models such as ResNet, combined with manual preprocessing steps supervised by radiology and orthopedic experts.系所名稱：國際醫學研究博士學位學程thesis