Skip to content

README.md

Latest commit

 

History

History
195 lines (141 loc) · 7.26 KB

README.md

File metadata and controls

195 lines (141 loc) · 7.26 KB

Parasite Image Classification System

End-to-end medical computer vision system for parasite microscopy image analysis, combining multi-class classification and an optional object detection branch for localization and crop-based workflows.

Parasite Samples

Overview

This project was built as an applied AI engineering system, not only a model training exercise. It is designed to support practical microscopy analysis where speed, repeatability, and clear evaluation are critical.

The initial project context was participation in the IEEE MANSB Victoris 2 Final Competition hosted on Kaggle, then extension into a broader end-to-end AI workflow.

The system addresses the problem of manual parasite identification by providing:

  • Structured image preprocessing and augmentation to improve robustness.
  • Multi-architecture classification experiments (Custom CNN and transfer learning models).
  • Optional YOLOv8-based detection pipeline for parasite localization and crop generation.
  • Evaluation artifacts for model comparison and operational decision-making.

Real-world use cases:

  • Laboratory screening support and triage assistance.
  • Research workflows for parasite species analysis at scale.
  • Educational and clinical prototyping for AI-assisted diagnostics.

Architecture

The platform follows a staged ML system design:

  1. Input Layer: Microscopy images from curated parasite datasets.
  2. Processing Layer: Data cleaning, resizing, augmentation, and dataset balancing.
  3. Modeling Layer: Classification training (Custom CNN, ResNet50, ResNet50V2, Xception) and optional YOLOv8 detection.
  4. Evaluation Layer: History curves, confusion matrices, AUC plots, and detection metrics.
  5. Inference Layer: Class prediction and optional object-level crop extraction.

System flow:

Microscopy Images -> Preprocessing/Augmentation -> Model Training -> Evaluation -> Inference Outputs

Architecture Diagram

The repository now includes the architecture diagram used for this staged ML flow:

Staged ML System Flow for Parasite Image Analysis

Features

  • Multi-class parasite image classification across 21 labeled classes.
  • Transfer learning support using ResNet50, ResNet50V2, and Xception.
  • Custom CNN baseline for architecture comparison.
  • Preprocessing and augmentation workflow to improve generalization.
  • YOLOv8 detection extension for object localization and crop extraction.
  • Experiment artifacts for each model family (history, AUC, confusion matrix, samples).
  • Notebook-first reproducible workflow from preparation to inference.

Dataset Sources and Label Taxonomy

Dataset sources used across the original workflow:

Primary classification label mapping (21 classes):

Label Description
0 Ascariasis
1 Babesia
2 Capillaria philippinensis
3 Enterobius vermicularis
4 Epidermophyton floccosum
5 Fasciolopsis buski
6 Hookworm egg
7 Hymenolepis diminuta
8 Hymenolepis nana
9 Leishmania
10 Leukocyte 400X
11 Leukocyte 1000X
12 Opisthorchis viverrine
13 Paragonimus spp
14 Plasmodium
15 Trichophyton rubrum (T. rubrum)
16 Taenia spp
17 Toxoplasma
18 Trichomonad
19 Trichuris trichiura
20 Trypanosome

Technical Highlights

  • Designed as a full pipeline from data preparation to inference, not a single notebook experiment.
  • Used transfer learning as a deliberate design choice to improve sample efficiency on specialized medical imagery.
  • Maintained both classification and detection capabilities to support multiple operational modes.
  • Preserved evaluation transparency through architecture-specific visual diagnostics stored in-repo.
  • Structured code and notebooks to support extension into API-based or cloud-based serving.

Model Details

  • Baseline model: Custom CNN for controlled comparison.
  • Transfer models: ResNet50, ResNet50V2, Xception.
  • Detection model: YOLOv8 (custom dataset training in parasite detection/TrainYolov8CustomDataset.ipynb).
  • Evaluation approach:
    • Classification: confusion matrix, AUC, training history, sample predictions.
    • Detection: precision, recall, mAP@50, mAP@50-95 tracked per epoch.

Tech Stack

  • Language: Python
  • Notebooks: Jupyter
  • Deep Learning: TensorFlow/Keras, PyTorch
  • Computer Vision: OpenCV, Ultralytics YOLOv8
  • Data and Analysis: NumPy, pandas, scikit-learn, Matplotlib
  • Experiment Tracking: TensorBoard logs (YOLO training runs)
  • Cloud Readiness: Microsoft Azure / Azure Machine Learning (integration-ready workflow)

Skills Demonstrated

  • Data preprocessing, feature engineering, and dataset balancing.
  • Computer vision and image processing with OpenCV.
  • Transfer learning and deep learning model development.
  • Classification modeling and predictive analytics.
  • Data analysis with NumPy and pandas.
  • Statistics-driven model evaluation and reporting.
  • API design mindset for downstream integration.
  • Cloud-aligned ML workflows (Azure and Azure ML readiness).

Getting Started

Prerequisites

  • Python 3.10+
  • pip
  • Jupyter Notebook or VS Code with Jupyter extension

Installation

git clone https://github.com/<your-username>/Parasite-Image-Classifier.git
cd Parasite-Image-Classifier
python -m venv .venv
.venv\Scripts\activate
pip install --upgrade pip
pip install jupyter numpy pandas matplotlib scikit-learn opencv-python tensorflow torch torchvision ultralytics

Run the Classification Workflow

Open notebooks in this sequence:

  1. 1 - Perpare Data.ipynb
  2. 2 - Data Anaylsis.ipynb
  3. 3 - Deep Learning Model.ipynb
  4. 4 - Transfer Learning.ipynb
  5. 5 - Parasite Image Classifier.ipynb
jupyter notebook

Run the Detection Workflow

  • Training notebook: parasite detection/TrainYolov8CustomDataset.ipynb
  • Inference notebook: parasite detection/Detection.ipynb
  • Detection utility class: parasite detection/Main.py
  • Trained weights examples:
    • parasite detection/train/80-epochs/weights/best.pt
    • parasite detection/train/200-epochs/weights/best.pt

Results

Classification Artifacts

The repository includes architecture-level evaluation outputs under:

  • assets/Custom CNN
  • assets/ResNet50
  • assets/ResNet50V2
  • assets/Xception

Each folder contains model history, confusion matrix, AUC, and sample output visuals for technical review.

Detection Benchmarks (YOLOv8)

Metrics extracted from tracked training logs:

Run Best mAP@50 Epoch Best mAP@50-95 Epoch Final Precision Final Recall Final mAP@50 Final mAP@50-95
80 epochs 0.97071 47 0.81999 47 0.87429 0.89445 0.93801 0.79261
200 epochs 0.99442 90 0.84442 90 0.93633 0.90635 0.93236 0.76508

Engineering interpretation:

  • Longer training improved peak mAP performance.
  • Final-epoch drop in mAP@50-95 on the 200-epoch run suggests monitoring for overfitting and selecting weights by validation best epoch, not final epoch.