Deep learning may be able to help specialists differentiate pancreatic diseases on MR images, according to new findings published in Diagnostic and Interventional Imaging.
The team behind the study, representing China’s Shanghai Institute of Medical Imaging, wanted to see if a convolutional neural network (CNN) could provide radiologists with clarity in the face of “the complicated imaging appearances and anatomical relationships of pancreatic diseases.”
Data from more than 500 patients who underwent TI-weighted, contrast-enhanced MR examinations were included in the study. Studies from nearly 400 patients were used as a training set, with the rest of the data being used as “internal and external validation sets.” The training and validation data was collected from different facilities. Training sets were then augmented by synthetic images from generative adversarial networks, and all synthetic images were reviewed by a team of two radiologists.
“The classification ability of a CNN model heavily depends on its architecture,” wrote authors X. Gao and X. Wang of China’s Shanghai Institute of Medical Imaging. ”In this study, we chose the classic InceptionV4 model because the performance of the other CNN models, such as InceptionV3 and GoogLeNet, whose architecture was similar to that of InceptionV4 network, had been validated in the medical fields. For example, InceptionV3 network had been used to detect retinal diseases on retinal fundus photographs. GoogLeNet had also been applied in the classification of pulmonary and pleural abnormalities on radiographs.”
When asked to separate findings into one of seven categories, the patch-level average accuracy of the InceptionV4 model was 71.56% for the internal validation set and 79.46% for the external validation set. The network’s area under the ROC curve (AUC) was 0.9204 for the internal validation set and 0.9451. The performance of an experienced radiologist was also tracked, and the specialist had an average accuracy and micro-averaging AUC of 82% and 0.8950, respectively, for the internal validation set. The average accuracy and micro-averaging AUC for the external validation set, meanwhile, were 83.93% and 0.9063, respectively.
Also, the Cohen’s kappa coefficients, which are used to measure prediction agreement, between the CNN and the radiologist were 0.8339 for the internal validation set and 0.8862 for the external validation set.
“Our study demonstrated that deep learning had the potential to differentiate between various pancreatic diseases on contrast-enhanced MR images,” the authors concluded.