Back to Projects
demoML/AImldata

Brain Tumor Detection (YOLOv8)

Medical imaging object detection system using YOLOv8 for brain tumor detection, achieving 0.905 mAP@50 with reproducible training pipelines.

2025-09 — 2026-02
0.905 mAP@50 accuracy
IoU 0.73 latency
View Source

Key Highlights

  • Achieved mAP@50 = 0.905, IoU = 0.73, pixel accuracy = 0.91
  • Diagnosed dataset-label mismatch and reframed the problem
  • Built reproducible training pipeline with COCO↔YOLO converters

Overview

A machine learning pipeline for detecting brain tumors in medical imaging using YOLOv8 object detection, with emphasis on reproducibility and proper problem framing.

Problem

Medical imaging datasets often have annotation issues. The original dataset was labeled for semantic segmentation but contained rectangular annotation artifacts that made it unsuitable for that task.

Solution

Diagnosed the dataset-label mismatch and reframed the task from semantic segmentation to object detection (bounding boxes), which better matched the actual annotation structure.

My Contributions

  • Identified dataset annotation issues and correctly reframed the ML problem
  • Built fully reproducible training and inference pipeline
  • Developed COCO↔YOLO format converters for data preprocessing
  • Created evaluation scripts and visual diagnostics
  • Trained YOLOv8 model on Apple Silicon (MPS backend)

    • Dataset & Evaluation

    Dataset: Medical brain MRI scans with tumor annotations

    Evaluation: Standard object detection metrics including mAP@50, IoU, and pixel accuracy

    Results: mAP@50 = 0.905, mean IoU = 0.73, pixel accuracy = 0.91

    Limitations

  • Model trained on specific MRI scan types; may not generalize to different imaging protocols
  • Requires careful preprocessing for new data
  • Bounding box detection doesn't provide precise tumor boundaries

    • Challenges & Tradeoffs

    Challenge: Original semantic segmentation approach produced poor results due to annotation artifacts.

    Solution: Recognized that rectangular annotations were better suited for object detection; switched to YOLOv8 bounding box approach, dramatically improving results.