Summary
Providing accurate and detailed radiographic analysis from medical images can be critical and often need re-evaluation. Implementing an automated system to review findings without disrupting the workflow helps to reduce this risk. Such systems offer an additional layer of redundancy in the diagnostic process, which can minimize errors, enhance diagnostic efficiency, and improve patient outcomes.
Abstract
Introduction
Radiologists play a key role in interpreting medical images, but subtle details might be missed, affecting patient care. Machine learning (ML) systems, such as the You Only Look Oncev8-Attention Model (YOLOv8-AM), can analyze large datasets to identify features like microfractures that may be overlooked by humans. This study aims to enhance the YOLOv8-AM fracture detection model, initially developed by Ruiyang, through targeted improvements.
Methods
The YOLOv8 architecture includes Backbone, Neck, Head, and Loss Function components. To improve detection accuracy, several modifications were made. Training epochs were increased to 100 to allow more learning time. Data augmentation techniques, such as rotation, scaling, and flipping, were applied to enhance generalization. Hyperparameters, including learning rate and batch size, were adjusted for optimal performance. Four attention models were employed, and the enhanced model was trained on the pediatric distal radius fracture dataset, GRAZPEDWRI-DX.
Results
The enhanced YOLOv8 models (iYOLOv8) showed significant improvements in precision and other metrics compared to the original model. The iYOLOv8 + GC model achieved the highest accuracy at 93.5% and an F1-score of 67%, with a mAP50 of 65.7% and 3.62 hours of training time, reflecting the benefits of the Global Context block. The iYOLOv8 + ECA model demonstrated strong performance with a precision of 92.7% versus 89.8% and an F1-score of 66% versus 64% in the original model, highlighting the effectiveness of Efficient Channel Attention. The iYOLOv8 + ResBlockCBAM model improved precision to 92.6% while maintaining a stable mAP50. Although advanced attention modules increased training times due to higher computational complexity, all models showed decreased training times compared to the original, with inference times remaining stable. The ResBlockCBAM variant had the lowest inference time of 2.3 ms
Discussion
and Significance/Clinical Relevance: The improved YOLOv8 models, particularly the iYOLOv8 + ECA and iYOLOv8 + GC versions, demonstrated notable advancements in precision and F1 scores, suggesting they offer more reliable fracture detection and reduced training times. These improvements enhance the accuracy of radiographic analysis, providing an additional layer of diagnostic redundancy that can minimize errors and streamline workflow. However, the increased computational complexity associated with advanced attention modules poses challenges for real-time applications where both speed and accuracy are essential. Implementing such automated systems could significantly benefit patient care by accelerating detection and decision-making processes, ultimately improving diagnostic efficiency and patient outcomes.