CNN Brain Tumor Classification Based on Masked Image Using U-Net Segmentation

Brain tumors are a serious issue due to late detection and potential misdiagnosis from varying sizes and locations. The current state-of-the-art in brain tumor detection uses deep learning models, while classification is often achieved through transfer learning. However, these approaches lack sufficient validation on larger datasets. This work improves this by using U-Net for segmentation, transfer learning for classification, and validating on a larger dataset to enhance the accuracy and generalizability. This study utilizes two datasets: Figshare Brain Tumor Dataset and the Br35H dataset. After preprocessing, a U-Net model detects tumor regions with notable 99 % accuracy. Transfer learning was used to classify tumors to glioma, meningioma, pituitary or no tumor using VGG16, VGG19, DenseNet121 and InceptionV3. VGG16 and DenseNet121 achieved highest accuracy around 94%. This approach enables more accurate detection and diagnosis, leading faster treatment and better patient outcomes.