Lung Nodule Classification using Learnt Texture Features on a Single Patient Population

Wednesday, Dec. 2 3:50PM - 4:00PM Location: S406B

Lyndsey C. Pickup, MEng, DPhil, Oxford, United Kingdom (Presenter) Employee, Mirada Medical Ltd
Aambika Talwar, MA, MBBCHIR, Oxford, United Kingdom (Abstract Co-Author) Nothing to Disclose
Shameema Stalin, Oxford, United Kingdom (Abstract Co-Author) Nothing to Disclose
Aymeric Larrue, PhD, Oxford, United Kingdom (Abstract Co-Author) Employee, Mirada Medical Ltd
Djamal Boukerroui, PhD, Oxford, United Kingdom (Abstract Co-Author) Employee, Mirada Medical Ltd
Mark J. Gooding, MENG, DPhil, Oxford, United Kingdom (Abstract Co-Author) Employee, Mirada Medical Ltd
Fergus V. Gleeson, MBBS, Oxford, United Kingdom (Abstract Co-Author) Consultant, Alliance Medical Limited; Consultant, Blue Earth Diagnostics Limited; Consultant, Polarean, Inc;
Timor Kadir, Oxford, United Kingdom (Abstract Co-Author) Employee, Mirada Medical Ltd

To validate the use of texture features and a machine learning approach to generate a "probability-of-malignancy" score for lung nodules.


A database with 705 distinct pulmonary nodules (PNs) was created with contrast CTs from 139 patients in a selected geographical region. All patients with reported PNs from Jan-Apr 2013 were included; those with unavailable scans or malignancy status (by histology or 2-year stable follow-up) were excluded. The dataset contained 328 benign nodules, 7 primary cancers, and 370 metastases. 522 image texture features in 2D/3D were extracted from each PN and its borders (contoured using Mirada XD, Mirada Medical Ltd). These included Haralick, Gabor and Laws features, fractal dimensions, plus combinations and difference features, with dimensionality reduction using principal component analysis. A greedy algorithm selected maximally discriminative features one by one, and mapped feature responses to malignancy probabilities using a Support Vector Regressor (LibSVM). For robust analysis, the dataset was partitioned into distinct thirds: one for training, one for cross-validation (setting SVR parameters, using a simplex method), and one for testing (reporting AUC). For each feature set, 100 different splits were evaluated, with the mean AUC on each split being compared. A leave-one-out validation result was also computed, for ease of comparison to other work. The work was repeated on a dataset excluding patients undergoing chemotherapy at the time of the scan, leaving 160 malignant and 230 benign nodules.


A mean AUC of 0.872 (std 0.020) was obtained by the feature set selected. The best single feature was the standard deviation of a Gabor filter response on the nodule boundary, and the peak mean AUC overall was obtained with 40 features. The leave-one-out AUC was 0.905, and this increase is to be expected because leave-one-out is less robust to overfitting than the three-fold approach. For the chemo-free population, the AUC was 0.942.


This texture feature model is successful at discriminating malignant and benign nodules over a large selection of nodules drawn from a single patient population. Future work should include more primary cancers.


Differentiating malignant and benign pulmonary nodules is a common clinical problem in which software may help support clinical decisions and guide patient management.