Skip to content

Latest commit

 

History

History
71 lines (54 loc) · 3 KB

README.md

File metadata and controls

71 lines (54 loc) · 3 KB

Blood Cell Classification with Deep Learning

Course: Artificial Neural Networks and Deep Learning (AN2DL)
Team Members: Denis Sanduleanu, Jonatan Sciaky, Alessio Caggiano, Mattia Repetti Project Group Name: Gan’s n Roses

📌 Overview

This project focuses on classifying images of blood cells into 8 categories using state-of-the-art Deep Learning techniques. Starting from a custom CNN and evolving through transfer learning and advanced augmentation strategies, we progressively improved our model's generalization and accuracy.

🧠 Problem

  • Dataset: 13,759 RGB images (96×96) across 8 blood cell classes.
  • Challenges:
    • Class imbalance.
    • Limited dataset size.
    • Generalization gap between local results and external test platform.

🛠️ Approach

We followed an incremental approach:

  1. Custom CNN – Initial attempt, later abandoned due to poor generalization.
  2. Transfer Learning – Used pre-trained ConvNeXt models (Small, Base, Large).
  3. Augmentation – Implemented advanced techniques like:
    • CutMix
    • MixUp
    • RandAugment
    • GridMask
    • CutOut
    • Custom masking
  4. Regularization & Optimization:
    • Dropout layers
    • L1/L2 regularization
    • Early stopping
    • Optimizers: Adam, Nadam, AdamW, Lion

📊 Results

Best performance achieved with ConvNeXtLarge, trained with the following setup:

  • Batch size: 256
  • Training set: 42,000 images (augmented)
  • Validation set: Augmented with RandAug (70%)
  • Fine-tuning: Last 25–40% of layers trainable
  • Optimizer: AdamW (lr=1e-3, weight_decay=1e-5)
  • Dropout: 35%
  • Regularization: L1L2 (1e-3)
  • EarlyStopping: Patience = 4 (fine-tuning phase)

Final accuracy on the platform: 83%

📦 Repository Contents

  • PreProcess.ipynb: Notebook for dataset cleaning, outlier removal, and preprocessing (e.g., masking).
  • balanced_augmentation.ipynb: Implements and tests various advanced data augmentation strategies including RandAug, CutMix, and MixUp to handle class imbalance and improve generalization.
  • custom.ipynb: A custom CNN architecture built from scratch, used as a baseline for comparison.
  • convnextbase.ipynb: Model training and evaluation using the ConvNeXt Base architecture with transfer learning.
  • ConvNeXtSmall.ipynb: Implementation and tuning of ConvNeXt Small for classification.
  • convNextlarge.ipynb: Final and best-performing model using ConvNeXt Large with extensive augmentation, fine-tuning, and regularization.
  • Report1.pdf: Full project report with methodology, results, analysis, and conclusions.

📚 References

🔍 Future Work

  • Implement model ensembling to further boost accuracy.
  • Extend classifier to other types of cell classification tasks.
  • Integrate explainability techniques for medical usability.