Deep Learning Models for BrainCancer Segmentation and Prognostic Analysis

Abstract

Deep Learning (DL) has introduced transformative potential in medical diagnostics, particularly in oncology, where complex tumor structures present significant challenges for treatment planning. The accurate segmentation of brain tumors from Magnetic Resonance Imaging (MRI) is a critical first step, and recent advancements in DL have greatly enhanced the efficiency and reliability of this process. By integrating these precise segmentation outputs with clinical and genetic data, prognostic analysis can be substantially improved, paving the way for more accurate tumor prediction and robust clinical decision support. This thesis aims to develop and evaluate a comprehensive pipeline for brain tumor analysis, from segmentationtosurvivalprediction.ItfocusesoncomparingadvancedDLarchitecturesforsegmentingbrain tumorsandexplorestheadaptationoflarge-scale,general-purposevisionmodelsforthisspecializedmedical task.Theultimategoalistofuseimagingfeatureswithclinicalandgeneticdatatoaccuratelyestimatepatient survival probability and stratify risk levels, thereby enhancing personalized treatment strategies. This work offers a thorough exploration of both established and cutting-edge DL models for brain tumor segmentation and prognosis. The first contribution is a comparative analysis of segmentation models, includingastandardU-Net,aDeepResUNet,andaVGG19-basedU-Net,alongsidefoundationalmodelslikea fine-tunedSegmentAnythingModel(SAM)andthespecializedMedicalSegmentAnythingModel(MedSAM). The results underscored the effectiveness of MedSAM, which demonstrated superior performance . The second contrirbution introduces a prognostic analysis pipeline that leverages features extracted from the segmentation masks. This pipeline proved highly effective, with the Support Vector Machine (SVM) model showing notable success in accurately identifying high-risk patients, while the XGBoost model demonstrated strong predictive power for classifying low-risk patients. Overall, these contributions validate a powerful, integrated approach for brain cancer analysis. The findingsconfirmthatleveragingbothCNN-basedarchitecturesandfine-tunedfoundationalmodels canyield highlyaccuratetumordelineations.Furthermore,thesuccessfulintegrationofthesesegmentationresultsinto aprognosticpipelinedemonstratesthecriticalroleofautomatedimageanalysisinfosteringpersonalizedand more effective clinical management of brain tumors.