Applying to TÜBİTAK 2209-A Undergraduate Research Support Program was my first serious encounter with academic research. My project idea was to develop a CNN system capable of multi-class classification between glioma, meningioma, pituitary tumor and healthy tissue from brain MRI images. A computer vision course I took at university had opened my eyes to this field, but I needed a systematic framework to go beyond existing literature — TÜBİTAK provided exactly that.
For the dataset, I used the open-source Kaggle Brain MRI dataset as my foundation, which contained nearly 3000 images with significant class imbalance. My augmentation strategy included horizontal/vertical flipping, ±15° rotation, brightness jitter and mixup. I normalized each class to 256×256 resolution and prepared it for transfer learning. I built the preprocessing pipeline with TensorFlow/Keras, achieving an efficient and reproducible workflow through the tf.data API.
For architecture selection, I fine-tuned ResNet-50, VGG-16 and EfficientNet-B3 with ImageNet weights. I froze base layers for the first 10 epochs while retraining the final classification head, then opened the full network with a low learning rate (1e-5) for the final 20 epochs. EfficientNet-B3 achieved the best performance with 93.4% validation accuracy. Analysis of the confusion matrix revealed that meningioma and pituitary tumors were most frequently confused — this stemmed from similar T1-contrast density distributions between the two tumor types.
What I learned most from the TÜBİTAK process was the written and oral dimension of research. Translating a technical study into academic language, defending it before a jury, and explaining your methodological choices clearly are skills just as important as the technical work itself. This experience has fundamentally shaped how I approach problem-solving and scientific communication.