This paper presents a reproducible 2D U-Net baseline for binary brain tumor segmentation trained on a publicly available MRI dataset (Nikhil Roxtomar, Kaggle, 3, 064 image–mask pairs). The work makes three explicit contributions beyond a standard implementation: (1) principled loss function selection using a manually implemented BCE + Dice combined loss, motivated by published hybrid loss comparisons; (2) a metric selection rationale that excludes pixel accuracy from model selection due to class imbalance inflation, using Dice coefficient as the primary criterion; and (3) a structured error taxonomy that categorises failure cases by attributed cause rather than visual appearance alone. The model achieves a mean Dice of 0. 5579 and IoU of 0. 4254 on non-empty test images (mean Dice 0. 4708 across all 64 test images). Training was conducted on consumer-grade hardware (NVIDIA GTX 1650 Ti, 4 GB VRAM) and completed in 58 minutes. The best checkpoint was selected at Epoch 11 based on validation Dice. Per-image Dice and IoU scores for all test images are available in the repository. v1. 1. 0 — Corrected evaluation protocol: 10 zero-Dice test cases correctly identified as genuine small-tumor prediction failures following ground truth mask inspection. All metrics now reported across full test set (n=64). Updated histogram figure. All code, model checkpoints, training scripts, and per-image evaluation results are publicly available at: https: //github. com/RaghavSarangan2003/brainₜumorₛegmentationᵤsingUNET
Raghav Sarangan Ramanujam (Fri,) studied this question.