Labeling and Semi-Supervised Learning on Medical Images (Brain MRI)

Python, Computer Vision, Deep Learning, Semi-Supervised Learning, Clustering, PyTorch, Scikit-learn, ResNet, Healthcare AI

Labeling and Semi-Supervised Learning for Brain Tumor Detection

Overview

This project implements a comprehensive analytical exploration of a largely unlabeled medical dataset, as part of an R&D project for automated brain tumor detection from MRI. The objective is to leverage a large volume of unlabeled images and a limited subset of expert labels to build a weak labeling pipeline followed by a semi-supervised classification model.

The project applies best practices of Computer Vision and semi-supervised learning to maximize the value of a partially labeled dataset, while evaluating the budgetary feasibility of scaling to several million images.

Objectives

• Explore and qualify a medical image dataset (resolution, channels, structure).
• Extract relevant visual features via a pre-trained model.
• Identify natural groupings through unsupervised clustering.
• Build a "weakly" labeled dataset from the clusters, without ever mixing it with expert labels.
• Implement and evaluate a semi-supervised approach, in comparison with classic supervised training.
• Formulate quantified recommendations for scaling (4M images, budget €5,000).

Demonstrated Skills

Data Exploration and Preparation

• Loading and visual exploration of an MRI dataset (resolution, color channel, file structure, metadata consistency).
• Preprocessing adapted to vision models: resizing, normalization.
• Methodical work on a reduced sample before scaling.

Feature Extraction by Transfer Learning

• Extraction of visual embeddings via a pre-trained ResNet-type model (frozen convolutional layers).
• Evaluation of several extraction layers to identify the most relevant representation.
• Constitution of a feature table usable for subsequent steps.

Clustering and Weak Labeling

• Dimensionality reduction (PCA, t-SNE) for visualization and interpretation of groupings.
• Application and comparison of several clustering algorithms (K-Means, DBSCAN).
• Generation of a "weakly" labeled dataset, kept strictly separate from the "strongly" labeled dataset.
• Validation of cluster relevance via the ARI score, calculated on the expert subset.

Semi-Supervised Learning

• Training a CNN model on the weakly labeled set, then continuing training on the strongly labeled set.
• Rigorous comparison of performance between purely supervised and semi-supervised approaches.
• Definition of metrics adapted to business risk (accuracy, F1-score, precision) and a clear "definition of done".
• Train/test split ensuring no data leakage during evaluation.

Feasibility Analysis and Recommendations

• Evaluation of AI labeling cost relative to available budget (€300 for the initial dataset).
• Argumented technical recommendations for scaling to 4 million images with a budget of €5,000, including feasibility conditions.
• Synthesis of results and trade-offs in the form of a presentation support for the project team.

Architecture

        MRI Dataset (majority unlabeled
         + expert labeled subset)
                      │
                      ▼
         Exploration & preprocessing
        (resolution, normalization, channels)
                      │
                      ▼
      Feature Extraction (Frozen ResNet)
                      │
                      ▼
        Dimensionality Reduction (PCA / t-SNE)
                      │
                      ▼
       Clustering (K-Means / DBSCAN)
                      │
        ┌─────────────┴─────────────┐
        ▼                           ▼
 "Weakly" Labeled            "Strongly" Labeled
 Set (clusters)              Set (experts)
        │                           │
        ▼                           │
 CNN Training                       │
 (weak labeling)                    │
        │                           │
        └─────────────┬─────────────┘
                      ��
       Semi-Supervised Training Continuation
                      │
                      ▼
     Supervised vs Semi-Supervised Comparison
                      │
                      ▼
   Scaling Recommendations

Technical Stack

Deep Learning & Computer Vision

• PyTorch / torchvision
• Pre-trained Models (ResNet)

Clustering & Dimensionality Reduction

• Scikit-learn (K-Means, DBSCAN, PCA)
• t-SNE

Data Manipulation & Visualization

• NumPy, pandas
• Matplotlib, seaborn, plotly
• OpenCV

Deliverables

• Jupyter Notebooks (.ipynb): feature extraction, clustering, semi-supervised approach
• Presentation Support (synthesis and recommendations)

Applied Best Practices

• Strict separation of weakly and strongly labeled datasets
• Freezing convolutional layers for stable transfer learning
• Cross-validation of clusters via a quantitative score (ARI) rather than purely visual interpretation
• A priori definition of evaluation metrics and a success criterion ("definition of done")
• Train/test split ensuring no data leakage
• Scaling recommendations based on a real budgetary constraint

Results

At the end of this project, the pipeline allows to:

• extract exploitable visual representations from a pre-trained model;
• generate reliable weak labeling on unlabeled images;
• train a semi-supervised model and quantify its performance gain compared to a classic supervised approach;
• determine the budgetary feasibility of scaling to 4 million images;
• provide a clear and actionable summary to a Data Science project team.

Acquired Skills

• Feature extraction by transfer learning (pre-trained CNNs)
• Unsupervised clustering and dimensionality reduction
• Construction and exploitation of weak labeling
• Semi-supervised learning and comparison of training protocols
• Definition of relevant metrics and success criteria
• Budgetary feasibility analysis and scaling recommendations
• Communication of technical results to a project team