Magnetic Resonance Imaging (MRI) stands as the gold standard for diagnosing a wide array of knee injuries, offering detailed insights into the joint’s intricate structures. However, the conventional interpretation of knee MRI scans is often a time-consuming process, heavily reliant on expert radiologists, and susceptible to both diagnostic errors and inter-observer variability. To address these challenges, the development of automated systems for knee MRI interpretation has emerged as a promising avenue, potentially streamlining workflows, prioritizing urgent cases, and aiding clinicians in achieving more accurate and timely diagnoses. Deep learning methodologies, renowned for their capacity to autonomously learn complex feature hierarchies, are particularly well-suited to model the intricate relationships between medical images and their clinical interpretations, paving the way for significant advancements in diagnostic accuracy and efficiency. This study delves into the creation of a deep learning model engineered for the detection of general knee abnormalities and specific diagnoses, notably anterior cruciate ligament (ACL) tears and meniscal tears, from knee MRI examinations. Furthermore, it investigates the tangible impact of providing model-generated predictions to clinical experts during their diagnostic interpretation process, aiming to quantify the potential for enhanced diagnostic performance.
Our research utilized a comprehensive dataset comprising 1,370 knee MRI exams, meticulously collected at Stanford University Medical Center between January 1, 2001, and December 31, 2012. The patient cohort presented a mean age of 38.0 years, with 569 (41.5%) female participants. To establish definitive diagnostic benchmarks, a consensus-based reference standard was derived from the majority opinion of three experienced musculoskeletal radiologists, applied to an internal validation set of 120 exams. We introduced MRNet, an innovative convolutional neural network meticulously designed for classifying MRI series. The predictions derived from three series per exam were subsequently integrated using logistic regression to yield a comprehensive diagnostic output. In performance evaluations on the internal validation set, MRNet demonstrated remarkable efficacy in detecting abnormalities, ACL tears, and meniscal tears, achieving Area Under the Receiver Operating Characteristic curve (AUC) values of 0.937 (95% CI 0.895, 0.980), 0.965 (95% CI 0.938, 0.993), and 0.847 (95% CI 0.780, 0.914) respectively. To further validate our findings, we incorporated a publicly accessible dataset encompassing 917 exams, featuring sagittal T1-weighted series and ACL injury labels, sourced from Clinical Hospital Centre Rijeka, Croatia. On an external validation set of 183 exams, MRNet, initially trained on Stanford sagittal T2-weighted series, exhibited a robust AUC of 0.824 (95% CI 0.757, 0.892) in ACL injury detection without any additional training. Notably, an MRNet model trained on the remaining external dataset achieved an even higher AUC of 0.911 (95% CI 0.864, 0.958). We rigorously assessed the specificity, sensitivity, and overall accuracy of nine clinical experts, including seven board-certified general radiologists and two orthopedic surgeons, on the internal validation set, both with and without the aid of model predictions. Statistical analysis employing a 2-sided Pearson’s chi-squared test, with adjustments for multiple comparisons, revealed no statistically significant disparities between the diagnostic performance of the model and that of unassisted general radiologists in abnormality detection. Interestingly, general radiologists demonstrated significantly superior sensitivity in ACL tear detection (p-value = 0.002; q-value = 0.019) and significantly higher specificity in meniscal tear detection (p-value = 0.003; q-value = 0.019). Further analysis using a 1-tailed t-test on the shift in performance metrics indicated that providing model predictions led to a significant enhancement in clinical experts’ specificity in ACL tear identification (p-value < 0.001; q-value < 0.001).
Our findings robustly demonstrate that our deep learning model possesses the capability to rapidly generate precise clinical pathology classifications from knee MRI exams, utilizing both internal and external datasets, thereby enhancing the accuracy of Diagnosis For knee injuries. Crucially, our results lend strong support to the premise that deep learning models can indeed augment the diagnostic performance of clinical experts in medical image interpretation. This suggests a future where AI-driven tools become integral in radiology workflows, improving diagnostic accuracy and potentially patient outcomes. Looking ahead, further research endeavors are essential to prospectively validate the model’s performance in real-world clinical settings and to meticulously evaluate its practical utility and integration within routine clinical practice. The potential of deep learning to transform knee injury diagnosis and medical image analysis is immense, promising a future of more efficient, accurate, and accessible healthcare.
