学术文献库

In the first step, a pre-trained model (YOLO) was used to detect all suspicious nod-ules. The YOLO model was re-trained using 397 CT images to detect the entire nodule in CT images. To maximize the sensitivity of the model, a confidence level (the probability threshold for object detection) of 0.3 was set for nodule detection in the first phase (ensuring the entire suspicious nodules are detected from the input CT images). The aim of the hierarchy model is to detect and classify the entire lung nodules (from CT images) with a low false-negative rate. Given the outcome of the first step, we proposed a 3D CNN classifier to analyze and classify the suspicious nodules detected by the YOLO model to achieve a nodule detection framework with a very low false-negative rate. This framework was evaluated using the LUNA 16 dataset, which consists of 888 CT images containing the location of 1186 nodules and 400000 non-nodules in the lung. Results: A large number of false positives were detected due to the low confidence level used in the YOLO model. Utilizing the 3D classifier, the accuracy of nodule detection was remarkably enhanced. The YOLO model detected 294 suspicious nodules (out of 321) when using a confidence level of 50%, wherein there were 107 false positives (187 true positives). By reducing the confidence level to 30%, 459 suspicious nodules were identified by the YOLO model, wherein 138 were false positives, and 321 were true positives. When the outcome of the YOLO model with a confidence level of 30% was fed into the 3D CNN classifier, a nodule detection accuracy of 98.4% and AUC of 98.9% were achieved Conclusion: The proposed framework resulted in a few false-negative and false-positives pre-dictions in nodule detection from CT images. The proposed approach would be helpful in detecting pulmonary nodules from CT images as a decision support tool.