Using Lightweight CNNs (U-Net Arch) for Brain Tumor Detection

We developed lightweight CNNs using a U-Net-inspired architecture for brain tumor detection from MRI scans, designed as a plug-and-play tool for hospitals to automate and standardize screening.

Context. Brain tumor incidence has risen, and MRI is the primary non-invasive imaging tool. Manual analysis is time-consuming and variable. We aimed to build an efficient, deployable model that could classify tumor presence and type from brain MRI (sagittal/x-z plane) with minimal misclassification—especially avoiding false negatives (missing a tumor).

Data & preprocessing. We used a Kaggle brain tumor dataset (~1,311 MRI images) with roughly balanced classes (Meningioma, Glioma, Pituitary, No Tumor). Data was split 70% train / 30% test. Preprocessing included grayscale conversion, Gaussian filtering to reduce noise, and edge detection (Sobel/Canny) to highlight tumor boundaries; images were resized via bilinear interpolation for efficient training.

Model. The pipeline uses convolutional layers with ReLU and max pooling, followed by dense layers and softmax for classification. The design prioritizes lightweight, interpretable blocks so hospitals can run it on typical hardware without heavy GPU infrastructure. The approach is aligned with U-Net-style thinking: focus on clear, hierarchical feature extraction and efficient use of spatial information.

Results. After training (e.g., ~10 epochs), we achieved high accuracy (~98.6% overall), with 100% recall on the no-tumor category and strong performance across tumor types (e.g., 99%+ for meningioma, glioma, pituitary). Weighted balanced accuracy was ~94.6%. Confusion matrices showed very few misclassifications, and errors were mostly between tumor subtypes rather than missing tumors entirely—critical for clinical safety.

Deployment. The goal is plug-and-play use in hospitals: ingest MRI scans, run the lightweight model, and get fast, consistent tumor detection to support (not replace) radiologist workflow. Future work includes full U-Net segmentation for precise tumor boundaries and support for multiple MRI planes (sagittal, coronal, axial) to improve robustness.