CytoCanine AI

AI-assisted detection and classification of canine round cell tumors (RCTs) from cytological images using deep learning.

This repository contains the complete implementation of the CytoCanine AI system, developed as part of the research study:

"Computer Vision Technology-Assisted Microscopic Detection of Round Cell Tumors in Dogs"

---

Project Overview

CytoCanine AI is a two-stage deep learning framework designed to assist in the automated detection and classification of canine round cell tumors (RCTs) from cytological microscopic images.

The pipeline integrates:

1. YOLOv8x – Tumor region detection (object detection)
2. ConvNeXt-Tiny (CNN) – Multi-class tumor classification
3. Grad-CAM – Explainable AI for visual interpretation

The system processes high-resolution cytology images, detects tumor regions, and classifies them into clinically relevant tumor categories.

---

Tumor Classes

The classification model predicts:

• Histiocytoma
• Lymphoma
• Mast Cell Tumor
• Transmissible Venereal Tumor (TVT)
• Negative (non-round cell tumor)

---

Dataset

The dataset used in this study is publicly available on Kaggle:

👉 https://www.kaggle.com/datasets/swatijaiswal46429/cytocanine-ai-dataset-yolo-cnn-cropped

It consists of three components:

1. YOLO Detection Dataset (Tumor-Detection-13)

• Annotated cytology images
• Bounding box labels for tumor regions
• Used for training YOLOv8

2. CNN Cropped Dataset (Cropped-Set)

• Generated from YOLO annotations
• Cropped tumor regions resized to 320×320
• Used for classification model training

3. Independent Test Dataset (276 images)

• Completely unseen dataset
• Used for final evaluation and reporting results

---

Live Demo

Interactive web application:

👉 https://huggingface.co/spaces/DeepBioSwati/CytoCanine_AI

Upload cytology images and get automated predictions.

---

Model Weights

Pretrained models are available on Hugging Face:

👉 https://huggingface.co/DeepBioSwati/CytoCanine_AI_models

Download and place inside:

models/

Available weights:

• yolov8x_best.pt – YOLO detection model
• convnext_tiny_final_earlystop.pth – CNN classification model

---

Repository Structure

CytoCanine_AI
│
├── training_code/
│   ├── cnn/
│   │   ├── train_convnext.py
│   │   ├── dataset_preparation.py
│   │   └── augmentation_pipeline.py
│   │
│   └── yolo/
│       ├── train_yolov8.py
│       ├── dataset_preparation.py
│       └── augmentation.py
│
├── inference/
│   ├── predict_pipeline.py
│   └── predict_yolo.py
│
├── evaluation/
│   ├── evaluate_pipeline_independent.py
│   ├── evaluate_cnn.py
│   ├── evaluate_yolo.py
│   ├── cnn_analysis_plots.py
│   └── yolo_analysis_plots.py
│
├── models/
│   └── placeholder.txt
│
├── app.py
├── README.md
├── requirements.txt
├── runtime.txt
└── packages.txt

---

Pipeline Workflow

The complete pipeline:

Input Cytology Image
        ↓
YOLOv8 → Detect tumor regions (bounding boxes)
        ↓
Patch Extraction (cropping)
        ↓
ConvNeXt-Tiny → Classify tumor type
        ↓
Aggregation Logic (patch-level → image-level)
        ↓
Final Prediction
        ↓
Grad-CAM Visualization (optional)

---

Installation

Clone the repository:

git clone https://github.com/Swati46429/CytoCanine_AI.git
cd CytoCanine_AI

Install dependencies:

pip install -r requirements.txt

---

Training

YOLO Training

python training_code/yolo/train_yolov8.py

Expected dataset format:

dataset/
 ├── train/images
 ├── train/labels
 ├── valid/images
 ├── valid/labels
 └── data.yaml

---

CNN Training

python training_code/cnn/train_convnext.py

Expected dataset format:

Cropped-Set/
 ├── train/
 ├── valid/
 └── test/

---

Evaluation

Full Pipeline (Recommended)

python evaluation/evaluate_pipeline_independent.py

Outputs:

• Accuracy
• Classification report
• Confusion matrix

---

Individual Evaluation

YOLO:

python evaluation/evaluate_yolo.py

CNN:

python evaluation/evaluate_cnn.py

---

Inference

Run full pipeline:

python inference/predict_pipeline.py --image path/to/image.jpg

Outputs:

• Final tumor prediction
• Confidence score
• YOLO detections
• Optional Grad-CAM visualization

---

Application Interface

Run the web app locally:

python app.py

---

Key Features

• Two-stage detection + classification pipeline
• Patch-based tumor classification
• Hybrid decision logic (CNN + YOLO override)
• Explainable AI using Grad-CAM
• Independent dataset evaluation
• Research-ready implementation

---

Citation

If you use this work, please cite:

@misc{CytoCanineAI_Dataset2025,
  title={Computer Vision-Assisted Technology Microscopic Detection of Canine Round Cell Tumors in Dogs},
  author={DeepBioSwati},
  year={2025},
  howpublished={\url{https://www.kaggle.com/datasets/swatijaiswal46429/cytocanine-ai-dataset-yolo-cnn-cropped}},
  note={Kaggle dataset for canine cytology image analysis}
}

License

This project is licensed under the MIT License.