Neural networks for nodal staging of non-small cell lung cancer with FDG PET and CT: importance of combining uptake values and sizes of nodes and primary tumor

Abstract

Purpose: To evaluate the effect of adding lymph node size to three previously explored artificial neural network (ANN) input parameters (primary tumor maximum standardized uptake value or tumor uptake, tumor size, and nodal uptake at N1, N2, and N3 stations) in the structure of the ANN. The goal was to allow the resulting ANN structure to relate lymph node uptake for size to primary tumor uptake for size in the determination of the status of nodes as human readers do.

Materials and methods: This prospective study was approved by the institutional review board, and informed consent was obtained from all participants. The authors developed a back-propagation ANN with one hidden layer and eight processing units. The data set used to train the network included node and tumor size and uptake from 133 patients with non-small cell lung cancer with surgically proved N status. Statistical analysis was performed with the paired t test.

Results: The ANN correctly predicted the N stage in 99.2% of cases, compared with 72.4% for the expert reader (P < .001). In categorization of N0 and N1 versus N2 and N3 disease, the ANN performed with 99.2% accuracy versus 92.2% for the expert reader (P < .001).

Conclusion: The ANN is 99.2% accurate in predicting surgical-pathologic nodal status with use of four fluorine 18 fluorodeoxyglucose (FDG) positron emission tomography (PET)/computed tomography (CT)-derived parameters. Malignant and benign inflammatory lymph nodes have overlapping appearances at FDG PET/CT but can be differentiated by ANNs when the crucial input of node size is used.

ANN training and testing protocol. In brief, the steps consist of randomly partitioning data set into two equal-sized groups called train and test, training the ANN on the former and testing it on the latter. At each of 100 iterations of this process, weights of the ANN that were obtained at training session are retained for the corresponding testing session. The output of the network is compared with the predictions of the expert PET reader and with true surgical-pathologic status at each iteration, and these accuracies are averaged over all iterations.