About NeuroScan AI
NeuroScan AI is a local, single-user tool for exploring automated brain tumor segmentation on multi-modal MRI. It was built by Hamza Mustafa and Wajid Ali as equal contributors, under the supervision of Ms. Saira Ishtiaq, as an end-to-end project spanning dataset preparation, model training, and an interactive viewer — with training run on Kaggle GPUs and everything else running entirely on the local machine.
The dataset
BraTS 2021 Task 1 provides pre-operative MRI scans from multiple institutions, each with four co-registered modalities — T1, T1-CE (contrast enhanced), T2, and FLAIR — and, for the training set, expert-annotated tumor sub-region labels.
Every voxel is labeled as background, necrotic/non-enhancing tumor core (NCR/NET, label 1), peritumoral edema (ED, label 2), or enhancing tumor (ET, label 4). Whole Tumor (WT) and Tumor Core (TC) are derived unions of these labels used for evaluation.
The model
A 3D U-Net takes all four modalities stacked as input channels and predicts three output channels for NCR, edema and enhancing tumor. Sigmoid scores are converted to one winning label per voxel; whole tumor and tumor core are derived unions, not separate model channels. Training happens on Kaggle GPU notebooks using the script in training/train_brats.py; the resulting checkpoint is copied to this machine and loaded by the local FastAPI backend.
The viewer
Scans are rendered client-side with NiiVue, a WebGL2 NIfTI viewer, so you can scroll through axial, coronal, and sagittal slices and toggle the predicted segmentation as a translucent color overlay directly on the raw scan.
3D anatomy credits
The landing-page brain is an anatomical illustration, not a patient scan or a segmentation result. Model by Z-Anatomy contributors and BodyParts3D / DBCLS, distributed through Brain Project under CC BY-SA 4.0. NeuroScan changes display materials and shows selected exterior structures; the bundled model is unmodified.