Sažetak | Tumori mozga, kao i svi ostali tumori nastaju nekontroliranom diobom stanica. Kako bih se ispravno planirala radioterapija, pratila progresija tumora te izabrala metoda liječenja, važno je ispravno ocrtavanje tumora. U ovom radu napravljen je model za automatsku segmentaciju GD pojačavajućeg tumora, peritumoralnog edema i nekrotične jezgre tumora bez povećanja. Za rješavanje ovog problema korištene su konvolucijske neuronske mreže, točnije U-Net arhitektura. Kod konvolucijskih neuronskih mreža protok informacija odvija se u samo jednom smjeru. U-Net arhitektura sastoji se od kodera, dekodera te mosta. Za izradu modela korišteni su Python, Keras, Google Colaboratory i ITK-snap. Za treniranje se koristio BraTS2020 skup podataka. Nakon predobrade i treniranja podataka dobiveni su sljedeći rezultati: Jaccard index iznosi 0.7123, Dice coefficient iznosi 0.7956, a iznos gubitka je 0.2151. Rezultati su relativno dobri, model se ne može koristiti kao zamjena za ručnu segmentaciju, ali se mogu koristiti za pomoć pri istoj. |
Sažetak (engleski) | Brain tumors, like all other tumors, arise from uncontrolled cell division. In order to properly plan radiotherapy, monitor tumor progression and choose a treatment method, it is important to correctly label the tumor. In this paper, a model for automatic segmentation of GD-enhancing tumor, peritumoral edema and necrotic non-enhancing tumor core. To solve this problem, convolutional neural networks, more precisely the U-Net architecture, were used. In convolutional neural networks, information flows in only one direction. The U-Net architecture consists of an encoder, a decoder and a bridge. Python, Keras, Google Collaboratory and ITK-snap were used to create the model. The BraTS2020 dataset was used for training. After data preprocessing and training, the following results were obtained: Jaccard index is 0.7123, Dice coefficient is 0.7956, and the amount of loss is 0.2151. The results are relatively good, the model cannot be used instead of manual segmentation, but it can be used to help with it.
Brain tumors, like all other tumors, arise from uncontrolled cell division. In order to properly plan radiotherapy, monitor tumor progression and choose a treatment method, it is important to correctly label the tumor. In this paper, a model for automatic segmentation of GD-enhancing tumor, peritumoral edema and necrotic non-enhancing tumor core. To solve this problem, convolutional neural networks, more precisely the U-Net architecture, were used. In convolutional neural networks, information flows in only one direction. The U-Net architecture consists of an encoder, a decoder and a bridge. Python, Keras, Google Collaboratory and ITK-snap were used to create the model. The BraTS2020 dataset was used for training. After data preprocessing and training, the following results were obtained: Jaccard index is 0.7123, Dice coefficient is 0.7956, and the amount of loss is 0.2151. The results are relatively good, the model cannot be used instead of manual segmentation, but it can be used to help with it. |