02058nas a2200229   4500000000100000000000100001008004100002260001200043653003900055653001100094653001600105653001600121100003100137700002100168700003600189245010000225856007900325300001000404490000600414520139400420022001401814       2021                            d  c09/202110aConvolutional Neural Network (CNN)10aFastai10aLung Cancer10aThoracic CT1 aSatheshkumar Kaliyugarasan1 aArvid Lundervold1 aAlexander Selvikvåg Lundervold00aPulmonary Nodule Classification in Lung Cancer from 3D Thoracic CT Scans Using fastai and MONAI  uhttps://www.ijimai.org/journal/sites/default/files/2021-08/ijimai6_7_8.pdf  a83-890 v63 aWe construct a convolutional neural network to classify pulmonary nodules as malignant or benign in the context of lung cancer. To construct and train our model, we use our novel extension of the fastai deep learning framework to 3D medical imaging tasks, combined with the MONAI deep learning library. We train and evaluate the model using a large, openly available data set of annotated thoracic CT scans. Our model achieves a nodule classification accuracy of 92.4% and a ROC AUC of 97% when compared to a “ground truth” based on multiple human raters subjective assessment of malignancy. We further evaluate our approach by predicting patient-level diagnoses of cancer, achieving a test set accuracy of 75%. This is higher than the 70% obtained by aggregating the human raters assessments. Class activation maps are applied to investigate the features used by our classifier, enabling a rudimentary level of explainability for what is otherwise close to “black box” predictions. As the classification of structures in chest CT scans is useful across a variety of diagnostic and prognostic tasks in radiology, our approach has broad applicability. As we aimed to construct a fully reproducible system that can be compared to new proposed methods and easily be adapted and extended, the full source code of our work is available at https://github.com/MMIV-ML/Lung-CT-fastai-2020.  a1989-1660