Part 2 - Image Recognition Methodology

As you may recall, one of the three options our customers will have to determine their correct breast pump flange size is:

Option 1 - Image Recognition

This method involves using AI tools to upload images of various breast sizes, allowing customers to take a picture of theirs and receive a size recommendation based on the closest cup size match. To achieve this, I needed three key components:

1. Images of all cup sizes

2. A training tool (I used Google’s Teachable Machine tool)

3. Test images

Ethical Considerations

Before diving into the data used, I want to address the ethics of this project. No personally identifiable data or sensitive images are collected, stored, or analyzed. Any images collected were sourced from publicly available databases and used strictly for training the model. I will not display any of the images used to create the classification system or for testing, as it is not necessary to demonstrate the effectiveness of the tool.

Data Sets

The dataset was sourced from a single website that contained images of breasts only—no faces or other body parts. These images were voluntarily submitted as part of an initiative to promote body positivity. I manually selected images that included size descriptions and used a subset of at least five images per cup size to test the Teachable Machine tool and create a training set.

For this test, I created training sets for A, B, C, and D cup sizes with the goal of evaluating:

• How many images were sufficient to create an accurate model

• Whether the system could differentiate between sizes that are close to one another (e.g., A vs. B)

• The overall effectiveness of the tool in detecting sizes

Outcome

Teachable Machine allows users to upload images for multiple classifications, train the model, and preview its accuracy before exporting it for external use. I used only 22 images across four size classes (A - D). Due to the small data set, I tested the model’s accuracy with a personal photo. Surprisingly, the model was fairly accurate despite the limited dataset—it was one size off, but the system classified me in both sizes, indicating some level of accuracy.

A larger sample size would improve accuracy. On Google’s AI Dev site, it states “When retraining a model with new data, you should aim to have approximately 100 data samples for each trained class. For example, if you are retraining an image classification model to recognize cats, dogs, and parrots, you should have around 100 images of cats, 100 images of dogs, and 100 images of parrots.” While a larger sample size would be ideal to test the accuracy, the difficulty in sourcing data for training remains a significant challenge.

Features For Option 1 - Image Recognition

AI-Powered Image Recognition - Uses machine learning to classify breast sizes based on uploaded photos.

Training Data - Dataset consists of publicly available breast images with size labels.

Classification Modeln - Built using Google’s Teachable Machine tool to categorize cup sizes.

Cold-Start Problem - Since the dataset is small, accuracy is limited, requiring more data over time.

User Experience & Implementation

We want the image recognition method to be simple and intuitive while ensuring privacy and accuracy. Here’s how users would interact with this option:

1. Uploading an Image

   • Users take a picture of their breast (without any identifying features).

   • The system processes the image and classifies it into the closest cup size category.

2. Machine Learning Model in Action

   • The AI model, trained using supervised learning, compares the uploaded image with known cup size classifications.

   • The system assigns a confidence score to the recommended size.

3. Results & Recommendations

   • The system provides the user with a flange size recommendation based on the detected breast size.

   • Additional guidance is provided if confidence in the classification is low.

Next Steps & Future Improvements

While this model would be an excellent way for our customers to determine their breast pump size, ethical considerations, limitations on data availability, and accuracy concerns prevent us from moving forward with this option for our first iteration. However, we plan to:

• Expand the dataset: Finding ethical, high-quality sources of breast images to improve accuracy and reduce the cold-start problem.

• Optimize the ML model: Implement more advanced deep learning techniques to improve classification accuracy.

• Improve User Privacy Measures: Explore on-device processing so no images are stored or sent to external servers.

• Validate with real users: Work with lactation consultants and testers to ensure recommendations align with actual sizing needs.

We are currently in discussions with a company that uses consumer-facing apps to measure breasts and areolas for plastic surgery evaluations. However, the user experience would be too cumbersome for busy moms, requiring them to take several pictures from different angles before receiving a measurement.

As a result, we will evaluate whether Option 2 (Favorite Brand’s Bra Size) provides a more user-friendly and scalable approach in the short term.

In my next post, I'll explore how Option 2 (Favorite Brand’s Bra Size) was developed and the key takeaways from our testing process. Stay tuned…