Abstract
1267
Purpose: Several methods of imaging analyses, including texture analysis and machine learning analysis, have been attempted to predict the prognosis of cancer patients. In particular, many studies have reported that several volumetric or texture features of baseline 18F-FDG PET are associated with the prognosis of cancer patients. In this study, we evaluated the performance of a deep learning algorithm using a convolutional neural network of baseline 18F-FDG PET to predict disease-free survival of soft tissue sarcoma.
Methods: We analyzed a total of forty-eight patients with soft tissue sarcoma, whose clinical and imaging data were from the Cancer Imaging Archive (TCIA: http://doi.org/10.7937/K9/TCIA.2015.7GO2GSKS). The baseline 18F-FDG PET images were used and divided into the training and test sets for predicting disease-free survival. The number of the training and test sets were 40 and 8, respectively. The deep learning algorithm was based on the convolutional neural network (CNN). The input data for CNNs were PET images, which were cropped manually including the primary tumor. The results were categorized either as recurrence or metastasis or as disease-free state. The sensitivity, specificity, and accuracy in predicting disease-free survival were used to evaluate deep learning performance. A 5-fold cross-validation method was employed.
Results: The sensitivity, specificity, and accuracy for predicting disease-free survival from the 5-fold cross-validation are as follows: for fold 0, 89.20%, 79.31%, and 90.47%; for fold 1, 94.20%, 75.29%, and 85.93%; for fold 2, 93.75%, 66.09%, and 81.66%; for fold 3, 91.07%, 84.48%, and 88.19%; for fold 4, 87.50%, 90.23%, and 88.69%, respectively. Cross-fold sensitivity was 92.68 ± 3.55%, specificity was 79.08 ± 9.17% and accuracy was 86.73 ± 3.10%.
Conclusions: The deep learning method using baseline 18F-FDG PET showed good performance in predicting disease-free survival in patients with soft tissue sarcoma. The deep learning method also has the advantage of convenience because there are few considerations for the researchers in the analysis compared to other imaging analysis methods. Therefore, the deep learning method can be a useful tool in predicting the prognosis of cancers.