ESTRO 2020 Abstract book

S316 ESTRO 2020

glioblastomas. Impact of normalization and discretization methods was evaluated based on a tumor grade classification task (accuracy measurement) using five machine learning algorithms (random forest, naïve bayes, logistic regression, support vector machine and neural networks Multi-layer Perception) (Figure 1).

PH-0533 MEGRE-NET-deep learning based automatic identification of fiducial markers in prostate RT MRI images

C. Jamtheim Gustafsson 1,2 , J. Swärd 3 , S. Ingi Adalbjörnsson 4 , A. Jacobsson 3 , O. Lars E 2

1 Skåne University Hospital, Haematology- Oncology and Radiation Physics, Lund, Sweden ; 2 Lund University, Medical Radiation Physics, Malmö, Sweden ; 3 Lund University, Centre for Mathematical Sciences- Mathematical Statistics, Lund, Sweden ; 4 Spectronic Medical AB, Spectronic Medical, Helsingborg, Sweden Purpose or Objective Identification of prostate gold fiducial markers can be challenging in MR images due to misclassifications from calcifications and post-biopsy fibrosis. Further, it is a time consuming task which can be associated with multiple operator related sources of uncertainties. Automated fiducial identification methods have therefore been suggested as an improvement. In multi-echo gradient echo (MEGRE) images a round signal void that expands rapidly around the fiducial with increasing echo time is utilized for fiducial identification, figure 1. The use of MEGRE images has previously been presented as promising solution with a 100 % manual detection accuracy. The aim of this work is therefore to develop a deep learning based method for automatic identification of gold fiducial markers in MR images intended for MRI-only prostate radiotherapy where CT is not available.

Results Normalization highly improves the robustness of first-order features and the performance of subsequent models on classification. For the T1w-gd sequence, the accuracy performance of the tumor grade classification model based on first-order features has increased from 0.68 to 0.83, 0.80, 0.82 respectively for the different standardization methods Nyul, WhiteStripe, Z-Score. Relative discretization makes unnecessary the use of normalization for textural features. The bin number has no major impact on classification performances when ranging from 16 to 128. There is a maximum variation of 10% in the percentages of robust features and a maximum variation of 9% in accuracy for bincount discretization in T2w-flair sequence for tumor grade classification based on texture (Figure 2).

Material and Methods MEGRE images from 150 prostate cancer patients with 3 inserted gold fiducials were acquired using a GE Discovery 750W 3T MRI system and post-processed with bias field correction. The fiducial center of mass (CoM) was identified in the MEGRE images and verified using CT. A sphere with an 8 mm radius was created around the CoM and defined the ground truth segmentation. A deep learning model was trained for semantic segmentation using the HighRes3DNet model in the NiftyNet framework with a 5-fold cross validation on two NVIDIA RTX 2080i Ti GPUs together with a relu activation function, L2 regularization and a CrossEntropy loss function. Images and ground truth segmentations were augmented by a one- axis random rotation before inputted to the network. The trained model was applied to unseen data from 39 MRI- only patients and 3D probability maps for fiducial location and segmentation were produced. Spatial smoothing of the probability map was performed by convolving it with a spherical kernel (8 mm radius) as a post-processing step. The final predicted segmentations were created by selecting only the 3 largest probability peaks (corresponding to 3 fiducials) and in those points insert 3 spherical objects with 8 mm radius, figure 2. The detection and geometric accuracy was assessed by comparing the predicted segmentation against the ground truth segmentation.

Conclusion A standardized pre-processing pipeline is proposed. For models based on first and second order features, we recommend to normalize images with the Z-Score method and to prefer an absolute discretization. For second-order features only, relative discretization can be used without prior normalization. In both cases, a number of bins between 16 and 128 is recommended. This study may paves the way for the multicentric development and validation of MR-based radiomic biomarkers.