Advancements in Deep Neural Network Architectures have enabledit to be used in the medical imaging and diagnostics domain, in termsof assisting doctors in making diagnoses. Although Neurodegenera-tive diseases such as Alzheimer’s disease remains a diagnosis basedon clinical grounds, there has been great development and recogni-tion that Deep Neural Networks can address the complexity and vol-ume of imaging data.The purpose of our Master Thesis is to address this problem of theclassification of brain PET scans with 18F-FDG radioactive tracer andtheir interpretability. For the classification, four pre-trained networkswere picked. These networks were trained on ADNI data and theirresults were compared to find the best model. To address the prob-lem of interpretability a hybrid interpretability model was developed;a model based on the combination of t-SNE plots and Class Activa-tion Maps. The hybrid model gave an insight into both the atomisticand holistic working of the network. In the end, the interpretabilitymodel was reviewed by a radiologist to check its effectiveness in thepractical world.Among the four models, the best performing model was the Incep-tionResNetV2 as it highlighted the hypometabolism activity in thebrain regions that are crucial to the diagnosis of Alzheimer’s Diseaseand explained the cluster formation of the correctly classified cases. Itwas concluded that the hybrid explainability model, provided betterinterpretability as compared to the presented individual interpretabil-ity methods.
Identifer | oai:union.ndltd.org:UPSALLA1/oai:DiVA.org:hh-46539 |
Date | January 2021 |
Creators | Pulicherla, Abhishek, Mujtaba Khan, Aiman |
Publisher | Högskolan i Halmstad, Akademin för informationsteknologi |
Source Sets | DiVA Archive at Upsalla University |
Language | English |
Detected Language | English |
Type | Student thesis, info:eu-repo/semantics/bachelorThesis, text |
Format | application/pdf |
Rights | info:eu-repo/semantics/openAccess |
Relation | Halmstad University Dissertations |
Page generated in 0.0016 seconds