Abstract WMP22: Synthetic Image Based Deep Learning of Stroke DWI Lesion Segmentation
| Autor: | Victor Schulze, Julian Maclaren, Johanna M. Ospel, Maarten G Lansberg, Soren Christensen, Kozerke Sebastian, Nino Scherrer, Christian Federau |
|---|---|
| Rok vydání: | 2020 |
| Předmět: | |
| Zdroj: | Stroke. 51 |
| ISSN: | 1524-4628 0039-2499 |
| DOI: | 10.1161/str.51.suppl_1.wmp22 |
| Popis: | Introduction: The application of deep learning to stroke image analysis (and medical images in general) faces two major challenges: first, it requires a large number of images to train, which is difficult to obtain. Second, the accurate outlining of infarcts is tedious, requires a high level of expertise, is subjective and error prone. The purpose of this work was to produce a large set of diffusion-weighted images (DWI) with perfectly defined realistic-appearing synthetic acute stroke lesions and to compare the segmentation performance of a deep neural network trained on these DWI scans with synthetic stroke lesions to a network trained on DWI scans of real stroke patients with stroke lesions manually outlined by neuroradiologists. Methods: 449 DWI scans with stroke lesions (72 ± 14y) and 2027 normal DWI scans (38 ± 24y) were coregistered, resampled, cropped to 96 x 80 x 40 voxels, normalized and divided into training/testing sets (80/20%). Stroke lesions were manually segmented by 3 neuroradiologists. 2000 synthetic 3D stroke DWI were produced by fusing thresholded (min 8%) signal increase of random DWI lesions to random coregistered normal DWI ( A ). A 3D U-Net (Tensorflow, depth 3, initial 64 feature maps doubled with each downsampling, bottleneck 3; hyperparameters optimized with cross-validation) was trained separately on 3 datasets: human-labeled real stroke cases (HL); 2000 synthetic cases (S); human-labeled real stroke cases + 2000 synthetic cases (HL+S). Results: The model trained on the human-labeled real stroke cases + 2000 synthetic cases (average dice coefficient between 300 and 600 epochs= 0.66±0.14) significantly outperformed the model trained on the 2000 synthetic cases only (0.60±0.14) and the model trained on human-labeled real stroke cases only (0.55±0.18; p-29 for all comparisons) ( B-C ). Conclusions: Deep learning segmentation of acute stroke lesions was significantly improved and was more stable by using synthetically generated images. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|