With an aging society, the need to automate time-consuming repetitive actions done by medicaldoctors to maximize their treatment ability is imminent. Automatic biomedical image segmentationalgorithms are set to play a key role in the healthcare of the future. Currently performed byradiologists, the time-consuming procedure consists of assigning areas on the image to correspondinganatomical structures. Novel automatic segmentation algorithms proposed in the literature can bedivided into atlas-based, methods using statistical shape knowledge and deep learning algorithms.Deep learning algorithms do not require complex preparation of the atlas or a priori knowledge aboutthe segmented shape. However, their performance is dependent on the training dataset size andquality. Employing the U-Net convolutional neural network architecture, the authors aim toovercome the bottleneck of a small-sized dataset with artificial data augmentation, creating newtraining samples using flipping and elastic deformation procedures. Algorithms' further increase ofefficiency was obtained by combining binary segmentation models – each model was trained tosegment one anatomical structure on the image. As most of the work in the field focuses on theintroducing novel neural networks' architectures to the field, the thorough description of the impactof these refinement steps sets the paper apart from the other publications in the field. The evaluationof the method utilized Dice's coefficient as a quantitative metric. The presented results show thedifferences between the model's coefficient values acquired on different magnetic resonancesequences used in the training process. Furthermore, data augmentation impact on segmentationaccuracy is showcased, as well as segmentation examples for visual inspection. The authors discussalso the practical usefulness of the algorithm, its limitations as well as future development plans.
展开▼
