Brain Tumor Classification & Segmentation

Deep Learning Pipeline – README

📌 Overview

This project implements a complete pipeline for brain tumor image analysis, including:

• Image classification
• Semantic segmentation
• COCO annotation processing
• Custom visualization and preprocessing tools

The notebook leverages a labeled dataset from Kaggle and applies modern deep-learning frameworks such as TensorFlow/Keras and scikit-learn, along with classical image-processing techniques.

---

📂 Dataset

• Source: Kaggle – Brain Tumor Image Dataset (Semantic Segmentation)
• Annotation type: COCO JSON (_annotations.coco.json)
• Includes bounding boxes and category labels for tumor regions.

Example categories as extracted from the annotation file:

• 0 – Non tumor
• 1, 2 – Tumor Categories (as defined by the author)

---

📦 Project Structure

The notebook uses the following components:

Imports

• General-purpose: numpy, pandas, json, random, matplotlib, seaborn

• Image processing: skimage, io

• DL frameworks: tensorflow, keras

• ML evaluation: accuracy_score, precision_score, f1_score, confusion_matrix, etc.

• Custom utilities from functions.py:

  • display_images_by_category
  • display_images_with_coco_annotations
  • visualize_annotation_mask
  • create_mask
  • extract_patches
  • segment_full_image

Reproducibility

All randomness is controlled via:

np.random.seed(42)
random.seed(42)
tf.random.set_seed(42)

---

🛠️ Notebook Workflow

1. Dataset Retrieval

The dataset is automatically downloaded using kagglehub. The script then loads COCO annotations:

with open('Dataset/train/_annotations.coco.json', 'r') as file:
    annotations = json.load(file)

2. Annotation Exploration

The notebook prints and inspects:

• Dataset metadata (version, licenses)
• Full image list
• Category definitions
• Bounding box structures

This is essential for understanding the segmentation task.

3. Custom Image Processing Tools

The custom functions allow:

• Visualizing images grouped by tumor category
• Displaying COCO bounding boxes
• Creating segmentation masks
• Extracting image patches for model training
• Segmenting full-size MRI images

4. Model Training & Evaluation

The notebook uses both ML and DL models:

• Logistic Regression (baseline classifier)

• CNN or UNet-like architectures (segmentation models)

• Evaluation through:

  • Accuracy
  • Precision
  • Recall
  • F1-score
  • Confusion matrix
  • Classification reports

---

📊 Visualizations

The notebook includes plots for:

• Samples from each tumor class
• COCO annotation overlays
• Generated segmentation masks
• Training curves (loss/accuracy)
• Model performance metrics

---

🚀 Goals of the Project

This notebook aims to explore and compare approaches for:

• Detecting brain tumor regions
• Classifying tumor types
• Segmenting tumor shapes and boundaries
• Understanding the dataset’s structure and annotations
• Creating a reusable training/visualization pipeline

---

👥 Authors

• Alessio Carnevale
• Manuel Cattoni
• Carlo Schillaci

---

📎 Notes

• This project requires a GPU-enabled environment for efficient training.
• Ensure the functions.py file is included in the working directory.
• COCO annotations must be kept in the correct relative folder structure (Dataset/).