Published
RSNA Annual Meeting (2022)
Authors
Taehee Kim1, Heejun Shin1, Juhyung Park1, Jongho Lee2, Hyungjin Kim3, and Dongmyung Shin1
Affiliations
1AI Engineering Division, Radisen Co. Ltd., Seoul, KOREA, REPUBLIC OF,
2Seoul National University, Seoul, KOREA, REPUBLIC OF,
3Seoul National University Hospital,
Seoul, KOREA, REPUBLIC OF.
Purpose
Recently, many deep learning-based nodule detection algorithms for chest X-ray radiographs (CXRs) have reported promising results. In real clinical environment, however, those algorithms have been reported to produce many false positives (FPs) caused by the similarity between hard negatives (e.g., nipple shadow, bone island, superimposed pulmonary vessels, etc.) and real nodules. Here, we propose a cascade-network to reduce these FPs in nodule detection.
Methods and Materials
NODE21 challenge dataset (node21.grand-challenge.org) was utilized (training: 934/2404; validation: 100/300; testing: 100/300 for nodule/non-nodule CXRs). For each CXR in the training dataset, a set of multi-resolution patches (resolution: 112×112, 280×280, and 448×448) were generated. Three multi-resolution patches were resized to 112×112 and concatenated along the channel for training. Patches were classified as positive (i.e., nodule patches) when a nodule locates on the center of the patch. For nodule detection, two neural networks were used sequentially where first network was to detect nodule candidates on an input CXR, and the second network was to further screen out hard negatives. Both networks had same architecture (EfficientNet-B0) and trained based on multiresolution
patches. The output of the two networks was the probability of nodule for the input patch. The difference between the networks was the composition of training dataset. For first network, all negative patches for training were the patches with no nodule (i.e., non-nodule patches). In second network, half of negative patches for training were carefully selected among the false positive outputs in the training data of first network to design second network more sensitive to the hard negatives. The detection performances were measured based on the area under the Free-Response Receiver Operating Characteristics (FAUC) and Refined Competition Performance Metric (R-CPM). Comparison was performed before and after ablating second network.
Results
The nodule detection performances of the cascade-network (FAUC:0.6033(95% CI: 0.5848-0.6217), R-CPM:0.5995(0.5778-0.6211)) outperformed those of the first network only (FAUC:0.4987(0.4667-0.5308), p<0.001, R-CPM:0.4967(0.4649-0.5285), p<0.001). When visually investigated, the outputs from the cascade-network showed a much smaller number of FPs.
Conclusions
In this study, we proposed a cascade-network for nodule detection to reduce FPs, dramatically improving nodule detection performance.
Clinical Relevance/Application
The proposed cascade-network with sequential training scheme would enable development of a nodule detection algorithm with high positive predictive value efficiently by using network-driven hard negatives.