🧬 Adaptive Unsharpening for Blood Cell Analysis

MATLAB and PyTorch implementations of adaptive focus-aware image enhancement and deep learning for blood-cell classification

Source code for the ICASSP 2021 paper
Acute Lymphoblastic Leukemia Detection Based on Adaptive Unsharpening and Deep Learning

---

🧠 Overview

This repository contains the original MATLAB implementation and a newer Python/PyTorch implementation of the adaptive unsharpening framework introduced in our ICASSP 2021 paper.

The method improves microscopic blood-cell images before classification by combining:

• focus-quality estimation;
• adaptive unsharp masking;
• VAR-PCANet / PCANet-based parameter tuning;
• pretrained deep CNNs and fine-tuning;
• classification and explainability analysis.

The original work was developed for Acute Lymphoblastic Leukemia (ALL) detection on ALL-IDB2. The PyTorch implementation additionally supports a modern GPU-oriented workflow and experiments on RAABIN-WBC.

---

✨ Original vs Adaptive Unsharpening

The adaptive preprocessing increases local detail while preserving the underlying cell morphology.

Example comparison on RAABIN-WBC and ALL-IDB2. Each original image is paired with the corresponding adaptively unsharpened version.

---

🔬 Method at a Glance

Microscopic blood-cell images
        │
        ▼
Focus-quality estimation
        │
        ▼
Adaptive unsharpening
        │
        ▼
VAR-PCANet / PCANet parameter tuning
        │
        ▼
Deep CNN feature extraction / fine-tuning
        │
        ▼
Classification
        │
        ▼
Blood-cell / leukemia prediction

The key idea is to adapt the amount of sharpening to the estimated focus quality of each image instead of applying a fixed sharpening strength to the entire dataset.

---

📁 Repository Structure

.
├── imgs/                       # README figures / example outputs
│
├── Matlab/                     # Original MATLAB implementation
│   ├── functions/
│   │   ├── functions_Classifiers/
│   │   ├── functions_DBProc/
│   │   ├── functions_ellipse/
│   │   ├── functions_FeatExtr/
│   │   ├── functions_Freq/
│   │   ├── functions_Gabor/
│   │   ├── functions_Kovesi/
│   │   ├── functions_Orient/
│   │   ├── functions_PCANet/
│   │   ├── functions_preProc/
│   │   └── functions_Reconstruct/
│   ├── libraries/
│   │   ├── lib_1Shot-MaxPol/
│   │   ├── lib_colorNorm/
│   │   ├── lib_FastCMeans/
│   │   ├── lib_FQPath/
│   │   └── lib_StainDeconv/
│   ├── params/
│   ├── steps/
│   └── util/
│
└── PyTorch/                    # GPU-oriented Python/PyTorch implementation
    ├── configs/
    ├── imgs/
    ├── scripts/
    ├── src/
    │   └── cnn_all/
    │       ├── classifiers/
    │       ├── cnn/
    │       ├── data/
    │       ├── focus/
    │       ├── legacy/
    │       ├── pcanet/
    │       └── pipeline/
    └── tests/

---

MATLAB Version

🚀 MATLAB Setup

The MATLAB implementation contains the original processing pipeline used for the ICASSP 2021 work.

Requirements

Typical components include:

• Deep Learning Toolbox
• Image Processing Toolbox
• Statistics and Machine Learning Toolbox
• Parallel Computing Toolbox

The repository also contains the external/supporting libraries used by the original pipeline.

Run

Open MATLAB from the repository root, move to:

Matlab/

and run:

launch_CNN_ALL

For the newer RAABIN-WBC adaptation, use the corresponding RAABIN launcher if present in your branch.

---

Python / PyTorch Version

⚡ GPU-Oriented Reimplementation

The PyTorch/ directory contains a modernized implementation designed for GPU execution.

The PyTorch version includes:

• automatic dataset/class discovery;
• focus-score caching;
• adaptive unsharpening;
• GPU PCANet / PCA operations;
• pretrained CNN feature extraction;
• CNN fine-tuning;
• k-NN classification;
• CMC and classification metrics;
• Grad-CAM;
• checkpointing and resume support;
• standalone export of the unsharpened dataset.

The shared preprocessing/focus-tuning stage is computed once and reused across CNN backbones, avoiding unnecessary repeated computation.

🐍 Installation

cd PyTorch

python3 -m venv .venv
source .venv/bin/activate

python -m pip install --upgrade pip setuptools wheel

Install a CUDA-enabled PyTorch build appropriate for your system, then install the project:

pip install -e .

Verify the installation:

cnn-all doctor

---

📂 Dataset Structure

RAABIN-WBC

PyTorch/
└── imgs/
    └── orig/
        └── Raabin-WBC/
            ├── Basophil/
            ├── Eosinophil/
            ├── Lymphocyte/
            ├── Monocyte/
            └── Neutrophil/

ALL-IDB2

ALL_IDB2/
├── 0/
└── 1/

with:

0 -> Healthy
1 -> Lymphoblast

Dataset paths and experiment settings are configured through YAML files in:

PyTorch/configs/

---

▶️ Running the PyTorch Pipeline

From the PyTorch/ directory:

source .venv/bin/activate

Run the full pipeline:

cnn-all run --config configs/raabin_wbc.yaml

Or run it phase-by-phase:

cnn-all prepare --config configs/raabin_wbc.yaml
cnn-all focus-cache --config configs/raabin_wbc.yaml
cnn-all shared --config configs/raabin_wbc.yaml
cnn-all networks --config configs/raabin_wbc.yaml
cnn-all export --config configs/raabin_wbc.yaml

For remote execution:

nohup cnn-all run --config configs/raabin_wbc.yaml > raabin_run.log 2>&1 &

Monitor with:

tail -f raabin_run.log

---

🧪 Supported CNN Backbones

The current PyTorch implementation supports:

• AlexNet
• VGG16
• VGG19
• ResNet18
• ResNet50
• ResNet101
• DenseNet201

The same original and adaptively unsharpened train/test splits can be reused across all backbones for a fair comparison.

---

📊 Expected Outputs

Output	Description
Classification results	Accuracy and related classification metrics
Original vs unsharp comparison	Evaluation on the two image versions
Focus statistics	Focus-quality values and selected thresholds
PCANet features	Learned PCA filters and feature representations
Trained CNNs	Fine-tuned network checkpoints
Grad-CAM maps	Visual explanation of CNN predictions
Logs	Experiment logs and configuration information
Exported database	Standalone adaptively unsharpened dataset

---

🧩 Included and Related Libraries

The MATLAB implementation includes or builds upon code and concepts from:

• PCANet
  T. Chan, K. Jia, S. Gao, J. Lu, Z. Zeng, and Y. Ma,
  PCANet: A Simple Deep Learning Baseline for Image Classification?
  IEEE Transactions on Image Processing, 2015.

• 1Shot-MaxPol
  M. S. Hosseini and K. N. Plataniotis,
  Convolutional Deblurring for Natural Imaging,
  IEEE Transactions on Image Processing, 2019.

• FQPath
  M. S. Hosseini et al.,
  Focus Quality Assessment of High-Throughput Whole Slide Imaging in Digital Pathology,
  IEEE Transactions on Medical Imaging, 2020.

• Fast N-D Grayscale Image Segmentation with c-/Fuzzy c-Means

• Stain Deconvolution / SCD_FastICA

• Comprehensive Colour Image Normalization
  G. Finlayson, B. Schiele, and J. L. Crowley, ECCV 1998.

---

📚 Paper

If you use this code, please cite:

@InProceedings{icassp21,
  author    = {A. Genovese and M. S. Hosseini and V. Piuri and K. N. Plataniotis and F. Scotti},
  title     = {Acute Lymphoblastic Leukemia Detection Based on Adaptive Unsharpening and Deep Learning},
  booktitle = {Proc. of the 2021 IEEE Int. Conf. on Acoustics, Speech, and Signal Processing (ICASSP 2021)},
  address   = {Toronto, ON, Canada},
  pages     = {1205--1209},
  month     = {June},
  day       = {6--11},
  year      = {2021},
  note      = {ISBN: 978-1-7281-7605-5. DOI: 10.1109/ICASSP39728.2021.9414362}
}

Paper:

https://ieeexplore.ieee.org/document/9414362

Project page:

https://iebil.di.unimi.it/cnnALL/index.htm

---

👥 Authors

• Angelo Genovese
  Department of Computer Science, Università degli Studi di Milano, Italy

• Mahdi S. Hosseini

• Vincenzo Piuri
  Department of Computer Science, Università degli Studi di Milano, Italy

• Konstantinos N. Plataniotis
  Department of Electrical and Computer Engineering, University of Toronto, Canada

• Fabio Scotti
  Department of Computer Science, Università degli Studi di Milano, Italy

---

📄 License

This project is released under the GNU General Public License v3.0.

See the LICENSE file for details.