Skip to content

AI · Computer Vision2026Graduation thesis

NeuroVision AI

Brain tumour classification from MRI scans: five deep learning architectures, measured data leakage and a Flutter app that presents the results.

Best accuracy
98.44%
MRI images
16,003
Classes
4
Architectures
5
Role
Research · ML · Mobile

Problem

Public MRI datasets often contain images from the same patient or near-identical copies. When these land in both the training and test sets, the model presents what it memorised as a "correct prediction" and the reported accuracy no longer reflects reality.

Solution

Every experiment shares the same deterministic data preparation. Exact copies are removed with SHA-256 and visually identical images with pHash; splitting is done by group and leakage checks are asserted in code. Horizontal flipping is not used for augmentation because of anatomical symmetry.

  • Four-class classification with per-class error analysis
  • Leakage measured by comparing accuracy before and after cleaning
  • A Flutter mobile app that presents the results

Process

  1. 01

    Dataset

    brain-tumor-dataset-v2 · 16,003 MRI images · 4 classes

  2. 02

    Duplicate cleaning

    Exact copies with SHA-256, visual duplicates with pHash

  3. 03

    Group-based split

    GroupShuffleSplit and leakage checks (assertions)

  4. 04

    Training

    Five architectures · Kaggle Tesla P100/T4 · SEED=42

  5. 05

    App

    Prediction and result screens in a Flutter mobile UI

Results

Test accuracy · cleaned dataset
ArchitectureAccuracy
ResNet5098.44%
DenseNet12195.63%
EfficientNetB0—
Custom CNN—

Findings

  • Before cleaning, the weakest class was meningioma; after pHash cleaning the weakest class became healthy (no tumour).
  • A limited but real data leak was confirmed.
  • The EarlyStopping callback carrying state across EfficientNetB0 runs was documented as a methodological finding.
  • Next step: segmentation on the BRISC2025 dataset using U-Net with an EfficientNetB0 encoder.

Screens

Technologies

  • Python
  • Deep Learning
  • Computer Vision
  • ResNet50
  • DenseNet121
  • EfficientNetB0
  • pHash
  • Kaggle GPU
  • Flutter
  • LaTeX

Advisor: Asst. Prof. Ali Rıza Gün · Bilecik Şeyh Edebali University