|
|
Utilization of convolutional neural networks for segmentation of mouse embryos cartilaginous tissue in micro-CT data
Poláková, Veronika ; Vičar, Tomáš (oponent) ; Chmelík, Jiří (vedoucí práce)
Automatic segmentation of the biological structures in micro-CT data is still a challenge since the object of interest (craniofacial cartilage in our case) is commonly not characterized by unique voxel intensity or sharp borders. In recent years, convolutional neural networks (CNNs) have become exceedingly popular in many areas of computer vision. Specifically, for biomedical image segmentation problems, U-Net architecture is widely used. However, in the case of micro-CT data, there is a question whether 3D CNN would not be more beneficial. The master thesis introduced CNN architecture based on V-Net as well as the methodology for data preprocessing and postprocessing. The baseline architecture was further optimized using advanced architectural modifications such as Atrous Spatial Pyramid Pooling (ASPP) module, Scaled Exponential Linear Unit (SELU) activation function, multi-output supervision and Dense blocks. For network learning, modern approaches were used including learning rate warmup or AdamW optimizer. Even though the 3D CNN do not outperform U-Net regarding the craniofacial cartilage segmentation, the optimization raises the median of Dice coefficient from 69.74 % to 80.01 %. Therefore, utilizing these advanced architectural modifications is highly encouraged as they can be easily added to any U-Net-like architecture and may remarkably improve the results.
|
|
|
Zlepšení kvality obrazu v rentgenové výpočetní mikrotomografii s velkým úhlovým krokem s využitím hlubokého učení
Šrámek, Vojtěch ; Šalplachta, Jakub (oponent) ; Zikmund, Tomáš (vedoucí práce)
Rentgenová výpočetní mikrotomografie představuje neinvazivní metodu, díky které jsme schopni zobrazit vnitřní strukturu objektů, proto se využívá v průmyslu i ve výzkumu. Doba měření rentgenových projekcí, které jsou potřebné pro tomografickou rekonstrukci obrazu objektu, se však může pohybovat v řádu desítek hodin. Jeden ze způsobů, jak zkrátit čas měření, spočívá ve zmenšení počtu naměřených projekcí, což ovšem negativním způsobem ovlivňuje kvalitu výsledné rekonstrukce obrazu. Pro zlepšení kvality rekonstruovaného obrazu je však možné aplikovat různé interpolační techniky. V této práci budou na data z laboratoře rentgenové mikro a nano výpočetní tomografie na CEITEC VUT a na data z veřejného zdroje aplikovány vybrané interpolační metody, které využívají hluboké učení, a bude vyhodnocena jejich úspěšnost.
|
|
|
Utilization of convolutional neural networks for segmentation of mouse embryos cartilaginous tissue in micro-CT data
Poláková, Veronika ; Vičar, Tomáš (oponent) ; Chmelík, Jiří (vedoucí práce)
Automatic segmentation of the biological structures in micro-CT data is still a challenge since the object of interest (craniofacial cartilage in our case) is commonly not characterized by unique voxel intensity or sharp borders. In recent years, convolutional neural networks (CNNs) have become exceedingly popular in many areas of computer vision. Specifically, for biomedical image segmentation problems, U-Net architecture is widely used. However, in the case of micro-CT data, there is a question whether 3D CNN would not be more beneficial. The master thesis introduced CNN architecture based on V-Net as well as the methodology for data preprocessing and postprocessing. The baseline architecture was further optimized using advanced architectural modifications such as Atrous Spatial Pyramid Pooling (ASPP) module, Scaled Exponential Linear Unit (SELU) activation function, multi-output supervision and Dense blocks. For network learning, modern approaches were used including learning rate warmup or AdamW optimizer. Even though the 3D CNN do not outperform U-Net regarding the craniofacial cartilage segmentation, the optimization raises the median of Dice coefficient from 69.74 % to 80.01 %. Therefore, utilizing these advanced architectural modifications is highly encouraged as they can be easily added to any U-Net-like architecture and may remarkably improve the results.
|
host ::
RSS.