RT Journal Article
SR Electronic
T1 PET Uptake Classification in Lymphoma and Lung Cancer using Deep Learning
JF Journal of Nuclear Medicine
JO J Nucl Med
FD Society of Nuclear Medicine
SP 325
OP 325
VO 59
IS supplement 1
A1 Ludovic Sibille
A1 Nemanja Avramovic
A1 Bruce Spottiswoode
A1 Michael Schaefers
A1 Sven Zuehlsdorff
A1 Jerome Declerck
YR 2018
UL http://jnm.snmjournals.org/content/59/supplement_1/325.abstract
AB 325Objectives: The interpretation of 18F-Fluorodeoxyglucose (FDG) PET/CT images is challenging given the sources of variability such as data acquisition, reconstruction methods and physiological tracer distribution. Recent advances in machine learning have enabled multi-parametric image classification with unprecedented accuracy. In this work, a convolutional neural network (CNN) was trained on PET/CT data to classify the malignancy of regions with increased uptake. Methods PET/CT images of 628 patients with known lung cancer or lymphoma were analyzed by two experienced nuclear physicians: each uptake region was annotated in anatomical location and physiological characterization (e.g. benign, malignant) and served as ground truth for the CNN. The algorithm incorporated image features such as multi-planar PET and CT reconstructions focused around each region of uptake to describe the region and its local context. A non-rigid registration of a maximum intensity projection (MIP) PET image to an anatomical atlas was used to characterize the anatomical location. The network comprised a series of alternating convolutional and pooling layers to extract imaging features while fully connected layers integrated the non-imaging features into a single network. The available data were divided into training (60%), validation (20%) and testing (20%) subsets. The validation was used to guide the algorithm development and to select the algorithm hyper-parameters, while the test set was exclusively used to assess the final algorithm performance parameters. Results A total of 12,959 PET regions of uptake were manually annotated from 302 lung cancer and 326 lymphoma cases resulting in 4,677 malignant lesions and 7,276 regions of physiological uptake. Characterization into benign or malignant classes using the CNN was shown to have an area under the ROC curve of 0.96 on the test subset demonstrating a good ability at discriminating benign versus malignant FDG uptake in this patient cohort. Furthermore, the MIP PET image proved to be useful to discriminate brown fat uptake and left-right symmetrical physiological uptake. Conclusions In this study we have demonstrated that the characterization of lesions in patients with lung cancer or lymphoma can be successfully performed using a convolutional neural network. While further image data and refinement of algorithms will be required for the application of such technology in clinical decision making, there are several potential use cases using the presented technology: reading workflow could potentially be simplified by preprocessing PET/CT to more efficiently derive clinically relevant parameters such as automatic parametric lesion description (e.g., total lesion glycolysis or total metabolic tumor burden), lesion differential diagnosis and pre-population of appropriate reporting fields. Finally, the same technique could be expanded onto other PET indications.