Background Deep learning neural networks have shown great promise for the automated analysis of complex medical images. Here we focus on the problem of discriminating malignant from benign lung nodules seen by computed tomography (CT). This is a major public health issue because only about 1 in 20 suspicious lung nodules that are seen on CT, many requiring invasive procedures, are actually cancerous.
Methods We constructed a multi-layer perceptron (MLP) neural network in which three layers of hidden neurons with 48, 24, and 12 nodes, respectively, connected to two output neurons corresponding to cancerous and benign. We also constructed two extended versions of the MLP by adding a preceding layer consisting of: (1) 64 3 × 3 parallel convolutional filters with learnable weights (CONV), and (2) 4 wavelet band-pass filters with pre-assigned weights (WAVE). MLP, CONV, and WAVE were each trained on 450 CT images of cancerous lung nodules and 845 images of benign nodules, using batch sizes of 30–100 images run over 30–150 epochs. MLP, CONV, and WAVE were tested on a separate set of 187 cancer and 114 benign images.
Results All nets achieved very similar classification accuracies of 74.3 ± 1.4% (combined mean). The training times for MLP and WAVE were within a mean of 1.5% of each other, but CONV took an average of 87 times longer to train as a result of its much greater number of free parameters.
Conclusion These findings suggest that the three networks extracted essentially the same amount of information pertaining to the presence or absence of malignancy, but that MLP and WAVE were much more efficient than CONV in this task.
This is a preview of subscription content, log in to check access.
This study was supported by NIH grants R01 HL-124052 and K23 HL133476.
Goodfellow I, Bengio Y, Courville A. Deep learning. Cambridge: MIT Press; 2016.Google Scholar
Siegel RL, Miller KD, Jemal A. Cancer statistics, 2015. CA Cancer J Clin. 2015;65(1):5–29.CrossRefGoogle Scholar
National Lung Screening Trial Research Team, Aberle DR, Adams AM, Berg CD, Black WC, Clapp JD, Fagerstrom RM, et al. Reduced lung-cancer mortality with low-dose computed tomographic screening. N Engl J Med. 2011;365(5):395–409.CrossRefGoogle Scholar
Kazerooni EA, Armstrong MR, Amorosa JK, Hernandez D, Liebscher LA, Nath H, et al. ACR CT accreditation program and the lung cancer screening program designation. J Am Coll Radiol. 2015;12(1):38–42.CrossRefPubMedGoogle Scholar
Hohberger LA, Schroeder DR, Bartholmai BJ, Yang P, Wendt CH, Bitterman PB, et al. Correlation of regional emphysema and lung cancer: a lung tissue research consortium-based study. J Thorac Oncol. 2014;9(5):639–45.CrossRefPubMedPubMedCentralGoogle Scholar
Hua KL, Hsu CH, Hidayati SC, Cheng WH, Chen YJ. Computer-aided classification of lung nodules on computed tomography images via deep learning technique. Onco Targets Ther. 2015;8:2015–22.PubMedPubMedCentralGoogle Scholar
Worley P, Prideaux D, Strasser R, Silagy C, Magarey J. Why we should teach undergraduate medical students in rural communities. Med J Aust. 2000;172:615–7.PubMedGoogle Scholar
Matsugu M, Mori K, Mitari Y, Kaneda Y. Subject independent facial expression recognition with robust face detection using a convolutional neural network. Neural Netw. 2003; 16(5–6):555–9.CrossRefPubMedGoogle Scholar