PT  - JOURNAL ARTICLE
AU  - Bradshaw, Tyler
AU  - Perk, Timothy
AU  - Chen, Song
AU  - Im, Hyung-Jun
AU  - Cho, Steve
AU  - Perlman, Scott
AU  - Jeraj, Robert
TI  - Deep learning for classification of benign and malignant bone lesions in [F-18]NaF PET/CT images.
DP  - 2018 May 01
TA  - Journal of Nuclear Medicine
PG  - 327--327
VI  - 59
IP  - supplement 1
4099  - http://jnm.snmjournals.org/content/59/supplement_1/327.short
4100  - http://jnm.snmjournals.org/content/59/supplement_1/327.full
SO  - J Nucl Med2018 May 01; 59
AB  - 327Objectives: Benign bone diseases and malignant bone lesions can have similar patterns of F-18 uptake in [F-18]NaF PET/CT imaging. This can make clinical evaluations challenging and inhibits the development of automated image analysis tools for bone scans. We evaluated the performance of a deep convolutional neural network for automatically differentiating benign and malignant lesions in NaF PET/CT images. Methods: Thirty-eight NaF PET/CT scans of patients with metastatic prostate cancer were evaluated by an experienced nuclear medicine physician. Evaluation consisted of identifying and scoring each bone lesion as (1) definitely benign, (2) likely benign, (3) equivocal, (4) likely malignant, and (5) definitely malignant. A subset of 14 scans was likewise evaluated by an additional 3 nuclear medicine physicians working independently, followed by a convened consensus among the physicians. Local 20 cm × 20 cm image patches centered on each lesion were extracted as inputs to a deep convolutional neural network, with each input patch consisting of 3 image channels: a maximum intensity projection (MIP) of the coronal PET, a MIP of the axial PET, and the axial CT. The network followed the VGG19 architecture, with 16 3×3 convolutional layers followed by 2 fully connected layers and a final softmax layer. The network’s weights were pre-trained using the ImageNet database of natural images, causing it to learn general image features. The network was then tuned using the lesion images and physician scores. The network’s accuracy (ACC), sensitivity (SEN), specificity (SPEC), and positive predictive values (PPV) were evaluated on an independent testing dataset comprising 30% of the data. Results: A total of 2,557 lesions were scored by one physician, 1,334 of which were also scored by the additional 3 physicians. When trained to differentiate benign (scores 1-2) and malignant (scores 4-5) lesions using the scores from a single physician, the network achieved an ACC, SEN, SPEC, and PPV of 0.82, 0.80, 0.84, and 0.88, respectively. When trained to predict the consensus scores of the physicians, the network’s performance worsened slightly to 0.86 (ACC), 0.73 (SEN), 0.86 (SPEC), and 0.82 (PPV). This reduction in performance is likely due to fewer training samples and imperfect agreement among physician scores. The network’s performance improved when trained to differentiate between definitely benign (score=1) and definitely malignant (score=5) lesions, with ACC, SEN, SPEC, and PPV of 0.88, 0.90, 0.85, and 0.90, respectively. Conclusions: Deep convolutional neural networks can be trained to accurately differentiate between benign and malignant bone lesions in NaF PET/CT images. Automatic classification of bone lesions may lead to improved clinical interpretation and facilitate the development of advanced image analysis tools for NaF PET/CT bone scans.