Magnetic resonance imaging diagnosis of knee injuries after skiing in adolescents under deep learning

To address the issues of subjectivity, low efficiency, and missed diagnoses of subtle injuries in magnetic resonance imaging (MRI) diagnosis of knee injuries after skiing in adolescents, a precise automatic diagnostic model was constructed to achieve simultaneous segmentation and classification of different types of injuries. A hybrid model combining U-Net + + and DenseNet121 was employed. U-Net + + performed pixel-level segmentation of injured areas, while DenseNet121 classified the injuries based on fused features. The model was trained using a dataset of 309 adolescent MRI scans through a joint loss function and transfer learning strategy. In the segmentation task, the average Dice coefficient (DC) was 0.89, and the intersection over union (IoU) was 0.82. The highest accuracy was achieved for meniscus tears (0.93, 0.87) and anterior cruciate ligament injuries (0.89, 0.82). Cartilage injuries (0.84, 0.77) showed a 6.4% improvement compared to the original U-Net. In the classification task, the average accuracy was 0.90, the F1-score was 0.91, and the area under the ROC curve (AUC) was 0.95. The recall rate for meniscus tears was 0.93, with a precision of 0.94, and the recall rate for cartilage injuries was 0.87. These results were significantly higher than those of SVM + handcrafted features (F1 = 0.77) and ResNet50 (F1 = 0.85) (P < 0.01). The model can efficiently and accurately perform automatic diagnosis of multiple injury types on MRI scans, outperforming traditional methods. It reduces the rate of missed and incorrect diagnoses, improves diagnostic consistency and efficiency, and holds value for clinical auxiliary applications.