目的以股骨转子间骨折数据集为研究对象,探究医用X线图像输入分辨率的改变对YOLO系列目标检测网络识别效果的影响。方法收集陆军特色医学中心2017-2022年间患转子间骨折患者的X线正位数据,按排除标准排除后共保留426例患者,847幅图像。...目的以股骨转子间骨折数据集为研究对象,探究医用X线图像输入分辨率的改变对YOLO系列目标检测网络识别效果的影响。方法收集陆军特色医学中心2017-2022年间患转子间骨折患者的X线正位数据,按排除标准排除后共保留426例患者,847幅图像。结合2018版Arbeitsgemeinschaft für Osteosynthesefragen/Orthopaedic Trauma Association(AO/OTA)分型指南和临床实际发病率,将转子间骨折重新整理划分为A1.2、A1.3、A2.2、A2.3、A3等5个亚组,并按8∶1∶1的比例划分为训练集(678幅)、验证集(84幅)和测试集(85幅),每次试验中保持严格一致。将YOLOX-Swin-Transformer、YOLOX、YOLOv5、YOLOv4目标识别网络图像输入大小分别设置为8种常用的分辨率,使用从头训练和迁移学习方法分别训练转子间骨折数据集,记录训练耗时,利用测试集进行模型测试并记录评价指标。采用SPSS20.0统计软件回归分析对训练耗时和均值平均精度(mean average precision,mAP)做曲线拟合,利用频数统计功能对各输入分辨率下评价指标评级为优的频数进行统计,以此判断最优区间。结果图像输入分辨率与各网络训练耗时呈正相关,所有P<0.05,线性回归分析有统计学意义。图像输入分辨率与网络mAP均值二次曲线拟合R^(2)=0.834,P=0.011,说明曲线拟合优度较好,回归分析有统计学意义。当图像输入分辨率在480×480、576×576、640×640范围内时,数据评级为优出现的频率最高(占42.86%)。结论训练耗时随分辨率的增加而增加。在不改变网络结构的前提下采用YOLO系列网络做医学图像识别的下游任务时,要取得最佳的识别效果,图像输入分辨率应控制在480×480、576×576、640×640范围之内。展开更多
Background: Intertrochanteric fractures are of intense interest globally. They are the most frequently operated fracture type, have the highest postoperative fatality rate of surgically treated fractures, and have bec...Background: Intertrochanteric fractures are of intense interest globally. They are the most frequently operated fracture type, have the highest postoperative fatality rate of surgically treated fractures, and have become a serious health resource issue due to the high cost of care required after injury. A number of problems exist when determining the best option for treatment for intertrochanteric fractures. The classification systems do not work well enough for preoperative planning and the reduction criteria have not been well defined. Methods: All patients who presented to Firoozgar Hospital, Tehran with intertrochanteric fracture, between March 20th 2013 and December 21st 2015, underwent DHS implementation, after 6-month follow-up period. Demographic data, preoperative radiographic fracture, bone quality typing (AO/OTA and Dorr classification), American Society of Anesthesiologists (ASA) score and evaluation of their correlation with the complication of fixation including nonunion and device failure were recorded and analyzed. Result: 101 patients including 54 males (53.5%) and 47 females (46.5%) underwent DHS implementation. Mean population age was 73.06 ± 16.36 years with an age range of 30 to 94 years;the most frequent age period was 76 - 85 years, injured by low-energy trauma mechanism and 23.8% patients were injured in a high-energy trauma. 56 patients completed the 6-month follow-up evaluation which ranged from 1 to 80 weeks. According to AO/OTA classification, the most common type among the patient population was A2, 27 patients (51.5%);while the most common group from Dorr classification turned to be group B (39.6%). Classification by ASA score revealed Class II to be most frequent among the patient population (56.4%). Nonunion was seen in eight of patients (14.3%). Three patients (5.4%) had device failure, two cases showed side plate breakage, and another one had screw cut out. There was no significant relationship between AO/OTA classification with the both complications i.e. the development of device failure (P = 0.85) and nonunion (P = 0.99). Non-significant correlation was found between Dorr classification with device failure (P = 0.06) and nonunion (P = 0.11). Conclusion: Regarding recent studies, more effective factor on the outcome is patient’s medical condition before surgery compared to the radiographic findings including AO/OTA and Dorr classification.展开更多
文摘目的以股骨转子间骨折数据集为研究对象,探究医用X线图像输入分辨率的改变对YOLO系列目标检测网络识别效果的影响。方法收集陆军特色医学中心2017-2022年间患转子间骨折患者的X线正位数据,按排除标准排除后共保留426例患者,847幅图像。结合2018版Arbeitsgemeinschaft für Osteosynthesefragen/Orthopaedic Trauma Association(AO/OTA)分型指南和临床实际发病率,将转子间骨折重新整理划分为A1.2、A1.3、A2.2、A2.3、A3等5个亚组,并按8∶1∶1的比例划分为训练集(678幅)、验证集(84幅)和测试集(85幅),每次试验中保持严格一致。将YOLOX-Swin-Transformer、YOLOX、YOLOv5、YOLOv4目标识别网络图像输入大小分别设置为8种常用的分辨率,使用从头训练和迁移学习方法分别训练转子间骨折数据集,记录训练耗时,利用测试集进行模型测试并记录评价指标。采用SPSS20.0统计软件回归分析对训练耗时和均值平均精度(mean average precision,mAP)做曲线拟合,利用频数统计功能对各输入分辨率下评价指标评级为优的频数进行统计,以此判断最优区间。结果图像输入分辨率与各网络训练耗时呈正相关,所有P<0.05,线性回归分析有统计学意义。图像输入分辨率与网络mAP均值二次曲线拟合R^(2)=0.834,P=0.011,说明曲线拟合优度较好,回归分析有统计学意义。当图像输入分辨率在480×480、576×576、640×640范围内时,数据评级为优出现的频率最高(占42.86%)。结论训练耗时随分辨率的增加而增加。在不改变网络结构的前提下采用YOLO系列网络做医学图像识别的下游任务时,要取得最佳的识别效果,图像输入分辨率应控制在480×480、576×576、640×640范围之内。
文摘Background: Intertrochanteric fractures are of intense interest globally. They are the most frequently operated fracture type, have the highest postoperative fatality rate of surgically treated fractures, and have become a serious health resource issue due to the high cost of care required after injury. A number of problems exist when determining the best option for treatment for intertrochanteric fractures. The classification systems do not work well enough for preoperative planning and the reduction criteria have not been well defined. Methods: All patients who presented to Firoozgar Hospital, Tehran with intertrochanteric fracture, between March 20th 2013 and December 21st 2015, underwent DHS implementation, after 6-month follow-up period. Demographic data, preoperative radiographic fracture, bone quality typing (AO/OTA and Dorr classification), American Society of Anesthesiologists (ASA) score and evaluation of their correlation with the complication of fixation including nonunion and device failure were recorded and analyzed. Result: 101 patients including 54 males (53.5%) and 47 females (46.5%) underwent DHS implementation. Mean population age was 73.06 ± 16.36 years with an age range of 30 to 94 years;the most frequent age period was 76 - 85 years, injured by low-energy trauma mechanism and 23.8% patients were injured in a high-energy trauma. 56 patients completed the 6-month follow-up evaluation which ranged from 1 to 80 weeks. According to AO/OTA classification, the most common type among the patient population was A2, 27 patients (51.5%);while the most common group from Dorr classification turned to be group B (39.6%). Classification by ASA score revealed Class II to be most frequent among the patient population (56.4%). Nonunion was seen in eight of patients (14.3%). Three patients (5.4%) had device failure, two cases showed side plate breakage, and another one had screw cut out. There was no significant relationship between AO/OTA classification with the both complications i.e. the development of device failure (P = 0.85) and nonunion (P = 0.99). Non-significant correlation was found between Dorr classification with device failure (P = 0.06) and nonunion (P = 0.11). Conclusion: Regarding recent studies, more effective factor on the outcome is patient’s medical condition before surgery compared to the radiographic findings including AO/OTA and Dorr classification.
