Neural networks for the analysis of small pulmonary nodules

Small pulmonary nodules can be readily detected by computed tomography (CT). The goal of this detection is to diagnose early lung cancer as the five year survival at this early stage is over 70% in contradistinction to the overall 5-year survival of around 10%. Critical to the efficacy of CT for early lung cancer detection is the ability to distinguish between benign and malignant nodules. We explored the usefulness of neural networks (NNs) to help in this differentiation.

CT images of 28 pulmonary nodules, 14 benign and 14 malignant, each having a diameter less than 3 cm were selected. All were sufficiently malignant in appearance to require needle biopsy and surgery. The statistical-multiple object detection and location system (S-MODALS) NN technique developed for automatic target recognition (ATR) was used to differentiate between these benign and malignant nodules.

S-MODALS was able to correctly identify all but three benign nodules. S-MODALS classified a nodule as malignant because it looked similar to other malignant nodules. It identified the most similar nodules to display them to the radiologist. The specific features of the nodule that determined its classification were also shown, so that S-MODALS is not simply a "black box" technique but gives insight into the NN diagnostics.

This initial evaluation of S-MODALS NNs using pulmonary nodules whose CT features were very suspicious for lung cancer demonstrated the potential to reduce the number of biopsies without missing malignant nodules. S-MODALS performed well, but additional optimization of the techniques specifically for CT images would further enhance its performance.