Hello, my name is Timothy Mitchell and I will be presenting work done which looked at using an artificial neural network to predict tumour occurrence in non-small cell lung cancer. I have no financial disclosures. Lung cancer is a leading cause of cancer mortality worldwide, largely treated with radiotherapy or a combination of chemoradiotherapy, as many cases are too advanced to treat. It would be useful to have a way to predict the response to those treatment strategies as they can be toxic to patients who at times have significant comorbidities. Artificial intelligence or AI is already being used in medicine and neural networks have been investigated as potential tools which can aid in a diagnostic setting. The aim of this study was to build an artificial neural network which could potentially predict tumour response in patients who receive radical radiotherapy for non-small cell lung cancer and to compare the performance of this ANN based model with classical logistic regression. 451 patients treated radical radiotherapy at the Western Park Cancer Centre were used for this study with independent and dependent variables investigated. The dependent variables included age, time between diagnosis and treatment, the lung PTB, the gender of the patients, tumour histology and the stage, while the dependent variable was tumour occurrence. Of the 451 patients, 361 were used for training the artificial neural network and 90 were used for testing once it was developed. The network consisted of 7 total layers, an input layer, 5 hidden layers and an output layer. The network also contained Leaky ReLU and SIGMOID activation functions compiled using the Adam Optimizer with a learning rate of 0.001. Analysis of the artificial neural network was repeated 3 times to determine the reproducibility. The metrics assessed in this study were derived from a confusion matrix, which is a 2x2 table outlining the true positives, true negatives, false positives and false negatives numbers. From there, the accuracy, precision, sensitivity, specificity and F1 score can be generated. A rock curve is also generated to measure how well the algorithm can distinguish between two classes. This table shows that the artificial neural network outperformed classical logistic regression when all metrics were assessed as well as scoring consistently above 70% in each metric. The diagram above demonstrates that the artificial neural network results were reproducible with similar F1 scores on each occasion. We assessed the reproducibility of the model using Monte Carlo cross-validation. The image above demonstrates that the rock AUC, or area under the curve, was 80%, showing excellent ability at distinguishing between the two classes, the depending and independent variables. So, our take-home message is that deep learning using an artificial neural network outperforms logistic regression as well as demonstrating that it can accurately predict tumor recurrence when assessing accuracy, precision, sensitivity and specificity as well as the F1 score. Further work will be looking at testing this algorithm in an independent dataset to see whether the network can perform independently.

artificial neural network

tumor occurrence

non-small cell lung cancer

radical radiotherapy

predict