ESTRO 2020 Abstract book

S996 ESTRO 2020

first channel input contained contrast enhanced slice with intensities clipped to the interval [375,425] HU. The second channel input contained slice with intensities normalized between ‐1000 and 400 HU. As an input to the third channel, histogram equalization was applied on the normalized axial slice. All the channels were individually rescaled to range [0,1]. The model was trained and optimized based on its predictions at multiple resolution levels (512, 256 and 128) which encouraged the network to predict correctly not only at the last layer, but also at deep layers with low resolution output. The performance of the model was evaluated on an external validation dataset from Maastricht University Medical Centre (MUMC) containing 170 patients with liver metastases of colorectal origin. The reference ground truth delineation was provided by a trained medical doctor. [1] P. Bilic et al. , “The Liver Tumor Segmentation Benchmark (LiTS),” 2019. [2] S. Xie, R. Girshick, P. Dollár, Z. Tu, K. He, and U. San Diego, “Aggregated Residual Transformations for Deep Neural Networks.”

Figure demonstrates excellent visual correspondence between a typical result (red) and the reference contour (blue) in axial, coronal, sagittal views (columns) for the investigated organs (rows). For all (6) test cases the mean DICE was 86.5, 86.4, 83.8, 81.5, the mean SFD was 1.1, 1.0, 0.8, 0.9, while the mean SFV‐2 was 88.1, 89.1, 94.3, 90.7 for PG‐left, PG‐right, SMG‐left, SMG‐right, respectively. Conclusion The proposed combination of 2D and 3D CNNs proved to be excellent in contouring small organs in T2 head MRI. The achieved accuracy was in line with the state‐of‐art (parotid gland, MRI, DICE 78‐91%) using only 25 training cases. In future, more training cases shall be used to increase the accuracy to 90‐95% DICE that is required by the clinical routine. PO-1710 A novel AI solution for auto-segmentation of multi-origin liver neoplasms A. Vaidyanathan 1,2 , W. Yousif MD 2 , A. Ibrahim MD 2,3 , B. Miraglio PhD 1 , R. Leijenaar PhD 1,2 , H. Woodruff PhD 2,3 , S. Walsh PhD 1 , P. Lambin MD PhD 2,3 1 Oncoradiomics SA, Research and Development, Liege, Belgium ; 2 The D-Lab and The M-Lab- GROW-School for Oncology and Developmental Biology- Maastricht University Medical Centre, Department of Precision Medicine, Maastricht, The Netherlands ; 3 GROW-School for Oncology and Developmental Biology- Maastricht University Medical Centre+, Department of Radiology and Nuclear Imaging, Maastricht, The Netherlands Purpose or Objective Liver cancer is one of the top three most deadly cancers in the world. Accurate, swift and reproducible liver lesion segmentation would benefit radiologists and oncologists with enhanced detection and diagnosis, treatment planning, clinical decision making, and monitoring. In this study, a novel convolutional neural network architecture trained to auto‐segment liver lesions is presented and evaluated. Material and Methods The LiTS‐Challenge dataset[1] containing 91 CT scans with different types of liver lesions(HCC, secondary liver tumours and metastasis) was used for training. A U‐Net based model adapted with three‐channel input and aggregated residual transformations[2] to extract important features was trained on the axial slices. The

Results The proposed model achieved an average dice score coefficient(DSC) of 0.72, an average sensitivity of 85% and an average False Positive Rate(FPR) <1% for their correct segmentation of metastasis. Total inference time for lesion segmentation took roughly 10s per case on a GTX 1070 GPU while it took around 15 min for manual segmentation. Figure 2 shows the test sample results and their corresponding statistics.

Conclusion We present a novel auto‐segmentation solution based on routine clinical imaging of metastatic liver cancer patients. The performance of the network is in line with human experts. The solution increases efficiency via swifter reporting times, especially in clinical and image analysis workflows. The solution also improves consistency, eliminating random human variation while admittingly introducing systematic bias that decreases