StankNet

StankNet: Deep Learning-Based Stool Classification using Transfer Learning StankNet is a deep learning model designed to classify stool samples according to the Purina Fecal Score Chart, powered by ResNet50 and transfer learning.

Overview

This tool serves pet owners in monitoring animal digestive health through automated, consistent scoring of stool samples. Our latest implementation shows significant improvements in classification accuracy through transfer learning.

Technical Highlights

• Built on ResNet50 architecture pre-trained on ImageNet
• Fine-tuned for specific stool classification tasks
• Demonstrated improved accuracy with strong diagonal pattern in confusion matrix
• Specialized in distinguishing between subtle differences in stool consistency

The Purina Fecal Score Chart (1-7)

• Score 1: Very hard and dry
• Score 2: Firm but not hard
• Score 3: Log-shaped, moist
• Score 4: Very moist but has shape
• Score 5: Very moist and barely has shape
• Score 6: Has texture but no shape
• Score 7: Watery, no texture

Model Performance

• Successfully differentiates between 7 different stool consistency classes
• Strongest performance in distinguishing middle-range consistencies (classes 2-5)
• Validated through confusion matrix analysis showing clear diagonal pattern
• Uses transfer learning to leverage ImageNet features while specializing in stool characteristics

Data Structure

Data collection starts inside of Google Drive then preprocessed and cleaned inside of an AWS S3 Bucket.

data/
    1/  # Very hard and dry samples
        image1.jpg
        image2.jpg
        ...
    2/  # Firm but not hard samples
        image1.jpg
        image2.jpg
        ...
    ...

Implementation Details

• ResNet50 backbone with custom classification head
• Transfer learning approach with frozen feature extraction layers
• Fine-tuned final layers for stool-specific feature detection
• Data augmentation and normalization for robust performance