Multi-pose-based convolutional neural network model for diagnosis of patients with central lumbar spinal stenosis.

Although the role of plain radiographs in diagnosing lumbar spinal stenosis (LSS) has declined in importance since the advent of magnetic resonance imaging (MRI), diagnostic ability of plain radiographs has improved dramatically when combined with deep learning. Previously, we developed a convolutional neural network (CNN) model using a radiograph for diagnosing LSS. In this study, we aimed to improve and generalize the performance of CNN models and overcome the limitation of the single-pose-based CNN (SP-CNN) model using multi-pose radiographs. Individuals with severe or no LSS, confirmed using MRI, were enrolled. Lateral radiographs of patients in three postures were collected. We developed a multi-pose-based CNN (MP-CNN) model using the encoders of the three SP-CNN model (extension, flexion, and neutral postures). We compared the validation results of the MP-CNN model using four algorithms pretrained with ImageNet. The MP-CNN model underwent additional internal and external validations to measure generalization performance. The ResNet50-based MP-CNN model achieved the largest area under the receiver operating characteristic curve (AUROC) of 91.4% (95% confidence interval [CI] 90.9-91.8%) for internal validation. The AUROC of the MP-CNN model were 91.3% (95% CI 90.7-91.9%) and 79.5% (95% CI 78.2-80.8%) for the extra-internal and external validation, respectively. The MP-CNN based heatmap offered a logical decision-making direction through optimized visualization. This model holds potential as a screening tool for LSS diagnosis, offering an explainable rationale for its prediction.

Investigation of optimal convolutional neural network conditions for thyroid ultrasound image analysis.

Neural network models have been used to analyze thyroid ultrasound (US) images and stratify malignancy risk of the thyroid nodules. We investigated the optimal neural network condition for thyroid US image analysis. We compared scratch and transfer learning models, performed stress tests in 10% increments, and compared the performance of three threshold values. All validation results indicated superiority of the transfer learning model over the scratch model. Stress test indicated that training the algorithm using 3902 images (70%) resulted in a performance which was similar to the full dataset (5575). Threshold 0.3 yielded high sensitivity (1% false negative) and low specificity (72% false positive), while 0.7 gave low sensitivity (22% false negative) and high specificity (23% false positive). Here we showed that transfer learning was more effective than scratch learning in terms of area under curve, sensitivity, specificity and negative/positive predictive value, that about 3900 images were minimally required to demonstrate an acceptable performance, and that algorithm performance can be customized according to the population characteristics by adjusting threshold value.