A Technical Overview: Developing a Fur Mask for Tumor Analysis

HT Vista by HT VET is an innovative screening tool designed to help veterinary teams differentiate between benign tumors and those requiring further investigation by analyzing their heat diffusion properties. The tool can assess both dermal and subcutaneous masses by gently heating the tumor and monitoring its heat diffusion with thermal and optical sensors. This process captures critical data that can aid in making informed diagnostic decisions.

In this post, we focus on a crucial pre-processing step in the analysis: the creation of a “Fur Mask.” Fur masks help identify areas with excessive fur that may obscure the thermal signals from the tumor and surrounding healthy tissue.

Why Extract the Fur Mask?

Fur naturally insulates heat, which can interfere with the analysis of tissue thermal properties by masking the thermal signal. To address this, veterinarians typically clip the fur around the tumor during the scanning process. However, tumors located in difficult-to-reach areas may have patches of longer fur that hinder accurate thermal measurements. Our Fur Mask algorithm detects these fur patches, enabling the system to ignore them and focus on the essential thermal data.

How Is It Done?

Given the uniqueness of our data, we developed a custom algorithm to address the challenges posed by fur. Our analysis revealed several key differences between fur and skin:

1. Initial Temperature Disparity: The initial temperature of fur is typically cooler than that of the surrounding skin.
2. Rapid Temperature Increase: Fur tends to heat up significantly faster, especially within the first few seconds of exposure to heat.
3. Distinct Textural Patterns: Fur often displays a unique stripe-like texture, identifiable in both thermal and optical imagery.
4. Color Influence: The rate at which fur and skin heat varies according to their color.

To tackle these challenges, we trained a neural network (NN) that assigns a probability score to each pixel, indicating whether it belongs to a patch of fur.

Neural Network Architecture

The neural network architecture we selected had to meet several key requirements:

• Effective with image inputs: The model needed to handle images as input and perform well on visual data.
• Pixel-level predictions: The input and output sizes should match, ensuring that the network makes predictions for each individual pixel (a task called image segmentation).
• Ability to recognize local textures: The network should be able to identify fine details and patterns within the image.
• Translation Invariance: The network should produce consistent results regardless of where the pixel is located in the image—whether it's at the center or along the edges.

To meet these criteria, we opted for the UNET architecture, which excels in image-based tasks. Its design makes it particularly well-suited for recognizing fine details and producing accurate pixel-level predictions, making it a top choice for our needs.

Image taken from – source

Input Data

The NN was trained using the following inputs:

Our training dataset consisted of approximately 500 scans, each manually labeled to ensure accurate training.

Training Results

The network was trained with the given input using common practices such as normalization, regularization and augmentations (image flips and rotations). The resulted trained network answered our needs and was able to detect long fur with high accuracy

Conclusion

The development of the Fur Mask is a critical step in our tumor analysis process, allowing for more precise diagnostics by filtering out noise caused by fur. By leveraging advanced neural network techniques and understanding the thermal dynamics of fur versus skin, we enhance the capabilities of our device, ultimately leading to better outcomes for veterinary patients.