No abstract

BackgroundAlthough deep learning systems (DLSs) have been developed to diagnose urine cytology, more evidence is required to prove if such systems can predict histopathology results as well.MethodsWe retrospectively retrieved urine cytology slides and matched histological results. High-power field panel images were annotated by a certified urological pathologist. A deep learning system was designed with a ResNet101 Faster R-CNN (faster region-based convolutional neural network). It was firstly built to spot cancer cells. Then, it was directly used to predict the likelihood of the presence of tissue malignancy.ResultsWe retrieved 441 positive cases and 395 negative cases. The development involved 387 positive cases, accounting for 2,668 labeled cells, to train the DLS to spot cancer cells. The DLS was then used to predict corresponding histopathology results. In an internal test set of 85 cases, the area under the curve (AUC) was 0.90 (95%CI 0.84–0.96), and the kappa score was 0.68 (95%CI 0.52–0.84), indicating substantial agreement. The F1 score was 0.56, sensitivity was 71% (95%CI 52%–85%), and specificity was 94% (95%CI 84%–98%). In an extra test set of 333 cases, the DLS achieved 0.25 false-positive cells per image. The AUC was 0.93 (95%CI 0.90–0.95), and the kappa score was 0.58 (95%CI 0.46–0.70) indicating moderate agreement. The F1 score was 0.66, sensitivity was 67% (95%CI 54%–78%), and specificity was 92% (95%CI 88%–95%).ConclusionsThe deep learning system could predict if there was malignancy using cytocentrifuged urine cytology images. The process was explainable since the prediction of malignancy was directly based on the abnormal cells selected by the model and can be verified by examining those candidate abnormal cells in each image. Thus, this DLS was not just a tool for pathologists in cytology diagnosis. It simultaneously provided novel histopathologic insights for urologists.

PurposeTo evaluate and compare the predictive performance of different deep learning models using gadolinium ethoxybenzyl diethylenetriamine pentaacetic acid (Gd-EOB-DTPA)-enhanced MRI in predicting microvascular invasion (MVI) in hepatocellular carcinoma.MethodsThe data of 233 patients with pathologically confirmed hepatocellular carcinoma (HCC) treated at our hospital from June 2016 to June 2021 were retrospectively analyzed. Three deep learning models were constructed based on three different delineate methods of the region of interest (ROI) using the Darwin Scientific Research Platform (Beijing Yizhun Intelligent Technology Co., Ltd., China). Manual segmentation of ROI was performed on the T1-weighted axial Hepatobiliary phase images. According to the ratio of 7:3, the samples were divided into a training set (N=163) and a validation set (N=70). The receiver operating characteristic (ROC) curve was used to evaluate the predictive performance of three models, and their sensitivity, specificity and accuracy were assessed.ResultsAmong 233 HCC patients, 109 were pathologically MVI positive, including 91 men and 18 women, with an average age of 58.20 ± 10.17 years; 124 patients were MVI negative, including 93 men and 31 women, with an average age of 58.26 ± 10.20 years. Among three deep learning models, 2D-expansion-DL model and 3D-DL model showed relatively good performance, the AUC value were 0.70 (P=0.003) (95% CI 0.57–0.82) and 0.72 (P<0.001) (95% CI 0.60–0.84), respectively. In the 2D-expansion-DL model, the accuracy, sensitivity and specificity were 0.7143, 0.739 and 0.688. In the 3D-DL model, the accuracy, sensitivity and specificity were 0.6714, 0.800 and 0.575, respectively. Compared with the 3D-DL model (based on 3D-ResNet), the 2D-DL model is smaller in scale and runs faster. The frames per second (FPS) for the 2D-DL model is 244.7566, which is much larger than that of the 3D-DL model (73.3374).ConclusionThe deep learning model based on Gd-EOB-DTPA-enhanced MRI could preoperatively evaluate MVI in HCC. Considering that the predictive performance of 2D-expansion-DL model was almost the same as the 3D-DL model and the former was relatively easy to implement, we prefer the 2D-expansion-DL model in practical research.

No abstract