Radiomics-net: Convolutional Neural Networks on FDG PET Images for predicting cancer treatment response

International audience324Objectives: The aim was i) to develop and validate three convolutional neural network architectures (CNN) for predicting response to cancer treatment in FDG PET imaging, and ii) to compare their performances with three random forest classifiers. Methods: We have developed an end-to-end 3D convolutional neural network (3D-CNN). We have also evaluated 2 others CNN architectures from the literature (Ypsilantis et al. PloS, 2015). The first one, called 1S-CNN, corresponds to an architecture where the input of the CNN corresponds to one slice. The process is repeated on each slice belonging to the tumor. Then, the model is evaluated by a majority vote. The second one, called 3S-CNN takes a triplet of adjacent slices as input. The CNN architectures were evaluated on a retrospective study database of initial FDG PET images of 97 patients with an esophageal cancer treated by chemo-radiotherapy. 56 patients responded 3 months after treatment. FDG positive tissues were segmented using a fixed threshold value of 40% of SUVmax. Images were spatially normalized with isotropic voxels (2x2x2 mm3) followed by an absolute intensity resampling (0.5 SUV/bin). For each CNN, the search for the best architecture was achieved using a validation procedure, by tuning hyper parameters, such as the number of layers, the number of feature maps and the size of filters among others. For comparison, a radiomic analysis was also conducted from 17 uncorrelated features (Spearman’s rank correlation coefficient < 0.8, p<0.05) using 3 random forest classifiers: without feature selection (WFS), with a selection strategy based on a genetic algorithm (GARF) and based on the forest’s importance coefficient (FIC) (Desbordes et al. PloS One 2017). A five-fold cross-validation was performed (57 patients for training, 20 for validation, 20 for test). The performances of the methods were evaluated after each cross-validation process including i) the accuracy (Acc) of the model corresponding to the percentage of patients correctly classified (responder vs. non responder) and ii) a receiver operating characteristic curve analysis computing the area under the curve (AUC), sensitivity (Se) and specificity (Sp). P-value < 0.05 was considered statistically significant. Results: The performances of the CNN architectures outperformed those found with the RF classifiers. The best results were found with 3D-CNN and 3S-CNN with comparable performances: Acc=0.72±0.08, AUC=0.70±0.04, Se=0.79±0.17 and Sp=0.62±0.21 (3D-CNN). 1S-CNN, seems to have lower performances (Acc=0.67±0.06, AUC=0.67±0.06), but the 1S-CNN ROC curve was not statistically significantly different from 3D-CNN (p=0.53) and 3S-CNN (p=0.48) ROC curves. From the RF classifiers, the best results were found with the GARF algorithm (Acc=0.68±0.08, Se=0.79±0.10, Sp=0.41±0.11, AUC=0.59±0.06). GARF ROC curve was not statistically significantly different from 1S-CNN (p=0.10) and 3S-CNN (p=0.058) ROC curves, while the 3D-CNN ROC curve gave better statistically significant results compared to GARF ROC curve (p=0.028). Conclusion: CNN architectures give very promising results for predicting cancer treatment response on initial PET tumor images. These results need to be confirmed with different type of cancers