Classification in Bone Scintigraphy Images Using Convolutional Neural Networks

The main goal of this thesis is to design and implement a convolutional neural network (CNN) to classify whether hotspots in bone scintigraphy images represent cancer metastases, caused by prostate cancer, or some other physiological process. A side task was included, where another CNN was designed and trained to classify whether a bone scintigraphy image image is acquired from the front of the patient (anterior) or from the back (posterior). The CNNs were implemented using the deep learning frameworks Deeplearning4J and Keras, and were trained on labelled data sets provided by Exini Diagnostics AB, Lund, Sweden. The performance of the CNNs were evaluated using various metrics commonly used in machine learning contexts, for instance accuracy and Receiver Operating Characteristics. The final trained CNN used for the classification of hotspots reached an accuracy of 89% on a test set consisting of images that had not been used for training. The corresponding CNN for the side task reached an accuracy of 99%. These results indicate that CNNs can work well for both classification problems.Skelettskintigrafi är en bildgivande undersökning liknande röntgen, där cancer metastaser och andra benskador visar sig som hotspots. Att identifiera dessa hotspots är en svår uppgift även för erfarna läkare. I detta arbete används faltningsbaserade neuronnät till att träna datorer att hjälpa läkare identifiera hotspots