Application of a Neural Network to Improve Nodal Staging Accuracy with 18F-FDG PET in Non-Small Cell Lung Cancer

摘要

id="p-1">We proposed to train a back-propagation artificial neural network (aNN) on a cohort of surgically proven non-small cell lung cancers (NSCLCs) and compare its accuracy with that of a trained 18F-FDG PET reader. We plan to show that an aNN trained on 18F-FDG PET- and CT-derived data is more accurate in predicting the true surgicopathologic nodal stage than a human reader. >Methods: One hundred thirty-three NSCLC patients with surgically proven N status treated at the University of Washington Medical Center or the Veterans Affairs Puget Sound Health Care System between February 1998 and September 2002 were used as inputs for the creation of an aNN. From CT of the thorax and 18F-FDG PET (neck to pelvis) performed before surgery, we extracted the primary tumor size and uptake (maximum pixel SUV [maxSUV]), normal lung and mediastinal uptake, and nodal uptake (maxSUV). Using the same 133 cases, the same output (surgical N status, N₀ to N₃), and the same software configuration settings, scenarios were created to assess which input parameters were most influential in creating an optimal aNN. To compute this optimal aNN, cases were split randomly 100 times into a training subset of 103 cases and a testing subset of 30 cases having the same proportion of N₀, N₁, N₂, and N₃ cases. N status predicted by the aNN was compared with the proven surgical N status to calculate the aNN accuracy. The N status readings from 18F-FDG PET were also compared with the surgical N status for the same cases to determine 18F-FDG PET accuracy. >Results: Statistical tests demonstrate that the best aNN accuracy is achieved by using N₁-N₂- N₃ nodal maxSUV divided by background uptake, the primary tumor size, and primary tumor maxSUV as inputs. The aNN correctly predicted the N stage in 87.3% of the testing cases compared with 73.5% for the 18F-FDG PET expert reader. Accuracy of the aNN increased to 94.8% (PET, 89.4%) when comparing N₀ + N₁ with N₂ or N₃ status and to 94.9% (PET, 91.9%) when comparing N₀ + N₁ with N₂ + N₃ status. >Conclusion: A back-propagation aNN can be trained to predict hilar and mediastinal nodal involvement with greater accuracy than an expert 18F-FDG PET reader. Such a tool could be used to improve clinical interpretations and for clinical training.