Detection of characteristic points on X-ray images using deep learning techniques

V tem magistrskem delu smo se ukvarjali s sistemom za detekcijo karakterističnih točk na slikovnih podatkih. Izdelali smo splošen sistem za avtomatsko zaznavanje karakterističnih točk, ki smo ga prilagodili za kefalometrične točke na rentgenskih slikah. Kot detektor smo uporabili obstoječo globoko nevronsko mrežo SegNet, ki je namenjena segmentiranju slik. To mrežo smo modificirali za iskanje karakterističnih točk. Novo mrežo smo poimenovali KeypointNet. Izdelali smo tudi sistem za označevanje točk na slikah. Detektor smo učili z učno množico. Na testni množici smo izvedli detekcijo in izmerili napako, ki jo definiramo kot evklidsko razdaljo med napovedano in označeno točko. Testirali smo tudi nabor krmilnih hiperparametrov pri zagonu učenja. Sistem smo testirali na množici 124 kefalometričnih slik velikosti 480 × 360 pikslov, in sicer za nabor 10 izbranih točk. Na vseh slikah smo označili te točke. Slike smo razdelili v učno in testno množico v razmerju približno 75 % in 25 %. Testirali smo 16 naborov hiperparametrov. Za vsak nabor smo izvedli 5 ponovitev učenja. Povprečna napaka v položaju točke na testni množici je bila 2,7 piksla. Testirali smo tudi vpliv dveh hiperparametrov za nadzor učenja. Testi so pokazali, da rahel odklon od priporočenih vrednosti za ta dva hiperparametra nima signifikantnega vpliva na končni rezultat. Dobljeni rezultati so zelo spodbudni. Razvili smo torej napreden sistem na osnovi globokega učenja, ki uspešno detektira karakteristične točke na slikah.In this master thesis we were dealing with a system for detecting characteristic points on image data. We have created a general system for automatic detection of characteristic points, which we have adapted for cephalometric points in X-ray images. As a detector we used the existing deep neural network SegNet, which is used for image segmentation. We modified this network to detect characteristic points. We named the new network KeypointNet. We also created a system for labeling points in the images. The detector was trained using a training set. The detection was performed on the test set and an error was calculated, which is defined as the euclidean distance between the predicted and the labeled point. We also tested the set of control hyperparameters when starting the learning process. The system was tested on a set of 124 cephalometric images of 480 × 360 pixels, for a set of 10 selected points. We\u27ve tagged these points on all the images. The images were divided into a learning and test set in a ratio of approximately 75% and 25%. We tested 16 sets of hyperparameters. For each set we performed 5 repetitions of learning. The average error in the point position on the test set was 2.7 pixels. We also tested the influence of two hyperparameters for learning control. The tests showed that a slight deviation from the recommended values for these two hyperparameters does not have a significant effect on the final result. The results obtained are very encouraging. Therefore we have developed an advanced system based on deep learning, which successfully detects characteristic points in the images