Pediatric Sarcoma Segmentation using Deep Learning

1208 Introduction: Pediatric sarcomas are a rare and complex cancer type consisting of several subtypes with different characteristics. Varying response to treatment within the same subtype has been observed clinically, calling for large systematic studies. Here, deep learning is used to implement an algorithm for automatic segmentation of pediatric sarcomas in PET/CT scans, with the final aim of paving the way for stratification studies. Methods: This retrospective study included 58 pediatric patients (32/26 m/f) age 6 mo. to 18 yo. with proven sarcomas and whole body 18F-FDG PET/CT. The main diagnoses were Rhabdomyosarcoma (RMS), Osteogenic sarcoma (OS) or Ewing sarcoma (ES). All images were anonymized. PET where acquired 60 min p.i of 3 MBq/kg body weight with a scan time of 3 min/bed. The CT were acquired in diagnostic quality with IV contrast. Additional 51 verified soft tissue sarcoma PET/CT data was obtained from The Cancer Imaging Archive (TCIA). These subjects had PET/CT with similar parameters. All tumors were delineated manually by a nuclear medicine specialist and served as gold standard for segmentation. A deep learning network that predict segmentation of sarcomas on the PET/CT data was trained using a model adapted from U-Net [1]. The model is build up by four encoding blocks, a base block, and four decoding blocks. For the convolution blocks maxpool is used for downscaling, ReLu as activation function, dropout and batch normalization is applied. Generalized Dice overlap was used as loss function. Before the final output, softmax activation is applied mapping each output pixel to a probability (tumor/background). The input is six channels; 3 consecutive 2D slices of PET data plus 3 consecutive 2D slices of CT data all with matrix size of 400 x 400. The output is two 400 x 400 images, one for each class (tumor/background) representing the center slice of the input slices. Data are presented to the network with random slice selection. During training an Adam optimizer is used with default values for β1= 0.9 and β2 = 0.999. The weights are initialized using He-initializer. The network is regularized by L2-weight regularization with a λ = 0.1 and by data augmentation using flip and rotation of the input. The model was evaluated by K-fold cross validation with K=5. Results: Hyperparameter search were performed and batch size=8, Learning rate= 1e-02, dropout rate as 0.1, 0.2, 0.3 at block level 1,2 and 3 respectively were found to be optimal. For tumor segmentation a voxel wise precision and sensitivity was found at 0.71/0.54 for Thorax, 0.71/0.39 for extremities, 0.52/0.38 for abdomen. For the head, neck and brain region the segmentation performed poorly mainly due to the high uptake in the brain, furthermore high FDG uptake in bladder and kidney also contributes to erroneous segmentations reflected by the precision and sensitivity scores. Discussion: A deep leaning network was trained to segment sarcomas on pediatric patients. Sarcomas are a very heterogenous class of tumors with large variations in size, shape and metabolism. This showed to be a challenging problem. Nevertheless, the network was able to accurately segment quite complex tumors (See figure, automatic segmentation in red outline) while in other cases it would fail. We attribute this to the limited number of training cases and hypothesize that an increasing number of training samples would improve the network. References: [1] Ronneberger et al, U-Net: Convolutional Networks for Biomedical Image Segmentation in Medical Image Computing and Computer-Assisted Intervention (MICCAI) 9351 (2015).