Introduction: To develop a generalizable and automatic Deep Learning model using 2D U-Net and LSTM algorithms to segment and discriminate two subtypes of papillary renal cell carcinoma (PRCC) through multiphase computed tomography (CT) images.
Methods: Total 71 patients underwent radical or partial nephrectomy or percutaneous renal tumor biopsy (RTB) and with pathologically proven PRCC were retrospectively included. Except for the test set (6 randomly selected cases each time), 85% and 15% of remaining data are for the model training and cross-validation, respectively. Following data pre-processing steps, automatic localization of region of interests (ROIs) and classification algorithm from multiphase CT images were implemented. A nested cross-validation was used for model optimization. The segmentation part was a 2D U-Net trained on the VISCERAL challenge dataset leveraging transfer learning, while the classification part was developed using spatial-frequency non-local convolutional LSTM network based on convolutional neural network (CNN). Further, the images were intensively discriminated by two urologists and the results of urologists were compared with that of the model.
Results: Among the 71 patients, 26 (36.6%) were proven to be PRCC type 1 and 45 (63.4%) were PRCC type 2. The trained 2D U-Net architecture automatic fixated most of its attention on the region of tumor masses in corticomedullary and nephrographic phases. After the training process of the CNN-based LSTM network, it performed for discrimination of two subtypes of PRCC with overall accuracy, sensitivity, and specificity of 78.5%, 66.7% and 81.5%, respectively, compared to those of the average performance of two experienced urologists with 66.7%, 88.5 and 68.9%, respectively (Table 1).
Conclusions: The CNN-based model on multiphase CT images can discriminate the two subtypes of PRCC with a competitive and superior performance and also shows up far faster than urologists which may be useful and practical in clinical work. Source of
Funding: This study was funded by the National Natural Science Foundation of China (81802535), China postdoctoral fund(223427), Nanjing Medical Science and technique Development Foundation (YKK 18064)