This thesis investigates the adaptation of the recurrent progressive fusion logic of RPF Net from multi source remote sensing to binary 2.5D multi sequence brain MRI patch classification. The study focuses on methodological evaluation rather than clinical diagnosis or tumor segmentation. Its main aim is to assess whether dual source representation, uncertainty aware trusted fusion, and DQN guided recurrent progressive fusion can support tumor/n on tumor patch classification under a strict 1% labelled training protocol. Braats derived MRI data were organized into two functional source roles: FLAIR as the first source and the combined T1, T1ce, and T2 sequences as the second source. A 2.5D input re presentation with K = 2 was used to include neighboring slice context. The non tumor label refers to the absence of tumor annotation within the sampled patch, not to a clinical healthy tissue category. The proposed pipeline was evaluated in three stages. S tage A learned a dual source 2.5D representation with initial average fusion. Stage B applied uncertainty aware trusted fusion using frozen Stage A outputs. Stage C used frozen Stage A and Stage B outputs within a DQN guided recurrent progressive fusion pr ocess. The final case disjoint held out evaluation contained 166,400 balanced patches from 1040 cases. Stage C, at the adopted threshold of 0.618, achieved the highest final accuracy and balanced accuracy among the three stages, both equal to 0.9728. It al so achieved the highest tumor class F1, reported as patch level classification Dice, of 0.9729 and the highest tumor class patch IOU of 0.9472, with 4,529 total errors. Stage B achieved the highest tumor sensitivity/recall, 0.9774. The improvement of Stage C over Stage C over Stage B was small, with a balancedStage B was small, with a balanced--accuracy gain of accuracy gain of 0.000210337 and 35 fewer errors out of 166,400 patches.0.000210337 and 35 fewer errors out of 166,400 patches. The findings show that RPF The findings show that RPF--NetNet--style recurrent progressive fusion can be style recurrent progressive fusion can be adapted methodologically to balanced adapted methodologically to balanced patchpatch--level MRI classification. level MRI classification. However, the Stage C gain should be interpreted as a small operatingHowever, the Stage C gain should be interpreted as a small operating--point refinement, not as evidence of clinical diagnosis, voxelpoint refinement, not as evidence of clinical diagnosis, voxel--wise wise segmentation, or tumorsegmentation, or tumor--boundary assessment.boundary assessment.
