DSHEAL_project_FS25_GaMo

Medical Image Analysis - Malaria Detection Using Blood Smear Images

Table of Content

• Task
• Short Project Description (Abstract)
• Structured Project Directory
• Dataset Details (Source & Statistics)
• Data Preprocessing
• Model Performance Summary
• Installation and Usage Instructions
• Evaluation
• Team Members and Contributions

---

Task

take a look at: https://github.zhaw.ch/ADLS-Digital-Health/DSHEAL-FS25/blob/main/project/assignment.md

---

Short Project Description (Abstract)

Malaria is a life-threatening disease that continues to impact millions globally, particularly in regions with limited access to healthcare resources. Manual diagnosis through microscopic analysis of blood smear images is both time-consuming and heavily reliant on expert knowledge. In this project, we developed a convolutional neural network (CNN) to automate the classification of malaria-infected cells using a publicly available dataset. The dataset contained labeled images of parasitized and uninfected cells, which were preprocessed through resizing, normalization, and augmentation techniques to improve model generalization. Our baseline model achieved a classification accuracy of approximately 95%, comparable to human expert performance.

To enhance model reliability, we explored hyperparameter tuning and interpretability methods, including threshold adjustment and LIME (Local Interpretable Model-agnostic Explanations). While threshold tuning successfully reduced false negatives, a critical factor in clinical settings, visual explanations from LIME revealed limited visible differences between correct and incorrect predictions, highlighting the ongoing challenges in model interpretability. Despite these limitations, our findings demonstrate that lightweight CNNs can serve as effective diagnostic support tools. Future work should focus on incorporating more diverse datasets, applying transfer learning, and validating the model in real-world clinical environments.

All code, trained models, and analysis results are included for full reproducibility.

Structured project directory

├── data/                       # Sample data or processing scripts (NOT the entire dataset)
│   ├── data_loader.py              # Data loading and transformation
│   ├── malaria_optuna.db           # Optuna study database
├── analysis/                   # Jupyter notebooks and scripts for analysis
│   ├── data_expd.ipynb             # Exploratory data analysis
│   ├── hyperparam.ipynb            # Hyperparameter analysis
│   ├── LIME_explanation.py         # LIME exploration script
│   ├── threshold_analysis.ipynb    # Threshold decision analysis
├── src/                        # Main scripts for training and evaluation
│   ├── evaluate.py                 # Evaluate the best model
│   ├── hyperparam_search.py        # Hyperparameter tuning with Optuna
│   ├── test_model.py               # Model testing script
│   ├── train_model_cv.py           # Model training with K-Fold cross-validation
│   ├── train_model.py              # Model training with predefined split
├── models/                     # Saved model weights and architectures
│   ├── best_malaria_model_*        # Best model from each run
│   ├── malaria_model_1.py          # Base model
│   ├── malaria_model_flex.py       # Flexible model, configurable via Optuna
├── report/                     # Final report and related documentation
│   ├── analysis_plots/             # Plots from exploratory and performance analysis
│   ├── performance_plots/          # Visualizations from Weights & Biases
├── requirements.txt            # List of Python dependencies
├── README.md                   # Instructions for using the project
├── results/                    # Evaluation results
│   ├── evaluation_best_*           # Evaluation outputs per model
├── wandb/                      # Weights & Biases log directory
├── env.yaml                    # Conda environment definition file
├── LICENSE                     # License file
└── README.md                   # Project overview and usage instructions

---

Dataset Details (Source & Statistics)

The dataset used in this project is publicly available on Kaggle:
https://www.kaggle.com/datasets/maestroalert/malaria-split-dataset/data

It contains cell images labeled as either Parasitized or Uninfected, already split into training, validation, and test sets.

File Type Summary

• Total image files: 27,558
• Format: .png only

Number of Images per Class

Set	Parasitized	Uninfected	Total
Train	11,024	11,024	22,048
Validation	1,378	1,377	2,755
Test	1,377	1,378	2,755
Total	13,779	13,779	27,558

Image Dimension Analysis

• Images vary in size, with most clustered around 128×128 pixels.
• Resizing to 128×128 was chosen for model input consistency.
• (See plots in analysis/data_expd.ipynb for more details.)

RGB Mean & Std (All Sets Combined)

Calculated across all train, validation, and test images:

• Mean: [0.5295, 0.4239, 0.4530]
• Std: [0.3323, 0.2682, 0.2821]

These values were used to normalize images during preprocessing.

---

Data Preprocessing

Before feeding the images into the neural network, a set of preprocessing and transformation steps is applied to ensure consistent input dimensions and improve model generalization.

Common Preprocessing Steps (All Sets)

• Resize: All images are resized to 128×128 pixels to standardize input dimensions.
• Normalization: Images are normalized using the RGB mean and standard deviation computed over the entire dataset:
  • mean = [0.529, 0.424, 0.453]
  • std = [0.332, 0.268, 0.282]
  • This scales pixel values to approximately the range [-1, 1].

Data Augmentation (Train Set Only)

To improve generalization and prevent overfitting, several augmentation strategies can be applied to the training set via the aug_level parameter:

Augmentation Level	Transformations Applied
none	Resize → ToTensor → Normalize
basic (default)	Resize → RandomHorizontalFlip → RandomRotation(+-15°) → ToTensor → Normalize
strong	Resize → RandomHorizontalFlip/VerticalFlip → RandomRotation(+-180°) → ColorJitter → Normalize

These options can be set via the get_train_transform() function.

Dataloader Setup

• Images are loaded using torchvision.datasets.ImageFolder(), which automatically assigns:
  • 0 = Parasitized (based on folder name)
  • 1 = Uninfected
• Data is then wrapped into PyTorch DataLoader objects with batching and shuffling enabled (for training).
• Validation and test sets are never augmented and use only resizing and normalization.

Data Path Handling

The dataset root path is stored in a separate data_path.txt file (not tracked by Git).
This path is automatically loaded via a utility function to keep code clean and portable.

---

Model Performance Summary

Evaluated Model: best_malaria_model_attmy8cd_epoch7
Augmentation Level: basic
Evaluation Set: Test Set (2,755 images)

Classification Report (Threshold = 0.50)

Class	Precision	Recall	F1-Score	Support
Uninfected	0.9640	0.9339	0.9487	1377
Parasitized	0.9360	0.9652	0.9503	1378
Accuracy			0.9495	2755
Macro avg	0.9500	0.9495	0.9495	2755
Weighted avg	0.9500	0.9495	0.9495	2755

Confusion Matrix (Threshold = 0.50):

• True Positives (TP, Parasitized correctly classified): 1330
• True Negatives (TN, Uninfected correctly classified): 1286
• False Positives (FP, Uninfected misclassified as Parasitized): 91
• False Negatives (FN, Parasitized missed): 48

---

Classification Report (Adjusted Threshold = 0.35)

Class	Precision	Recall	F1-Score	Support
Uninfected	0.9746	0.9187	0.9458	1377
Parasitized	0.9231	0.9761	0.9489	1378
Accuracy			0.9474	2755
Macro avg	0.9489	0.9474	0.9473	2755
Weighted avg	0.9488	0.9474	0.9473	2755

Confusion Matrix (Threshold = 0.35):

• True Positives (TP): 1345
• True Negatives (TN): 1265
• False Positives (FP): 112
• False Negatives (FN): 33

---

Interpretation

Adjusting the classification threshold from 0.50 to 0.35 significantly reduced the number of false negatives (from 48 to 33), which is crucial for medical diagnostics like malaria detection. Although this led to a slight increase in false positives (from 91 to 112), the overall accuracy remained high and nearly unchanged (~94.7%).

This trade-off aligns the model’s behavior with real-world priorities, where missing a malaria case is considered more critical than a false alarm.

---

Installation and Usage Instructions

1. Clone the Repository

Clone this repository and navigate into the project folder:

git clone https://github.zhaw.ch/mockand1/DSHEAL_project_FS25_GaMo.git
cd DSHEAL_project_FS25_GaMo

2. Environment Setup (via Conda)

Create and activate the Conda environment using the env.yaml file:

conda env create --file env.yaml
conda activate DSHEAL_proj1_GaMo

3. Data Download and Setup

1. Download the dataset from Kaggle:
   https://www.kaggle.com/datasets/maestroalert/malaria-split-dataset/data

2. Extract the dataset so that the folder structure looks like this:

your_data_path/data/
├── train/
│   ├── Parasitized/
│   └── Uninfected/
├── val/
│   ├── Parasitized/
│   └── Uninfected/
└── test/
    ├── Parasitized/
    └── Uninfected/

Make sure the folders train, val, and test are directly inside the data/ directory.
You do not need to rename or rearrange the subfolders — they are already correctly labeled.

3. Create a file named data_path.txt in the root of the cloned repository (DSHEAL_project_FS25_GaMo) and insert your local absolute path to the data/ directory.

Example content of data_path.txt (Windows):

C:\ZHAW_local\project_work\data

---

Evaluation

To evaluate the best model (or alternatively test all models in the models/ directory), run:

python src\evaluate.py --data_path "your_data_path/data"

Replace "your_data_path/data" with the actual path to your dataset.

---

Team members and contributions

Mike Gasser
ZHAW School of Engineering
gassemik@students.zhaw.ch

Andres Mock
ZHAW School of Engineering
mockand1@students.zhaw.ch