Using BreastMNIST: A Simpler Dataset for Deep Learning

BreastMNIST, a subset of the MedMNIST dataset, is a publicly available binary classification dataset designed for medical image analysis. Binary classification plays a critical role in medical imaging as it simplifies complex diagnostic tasks, allowing models to focus on identifying key differences between normal and pathological conditions. This specificity is particularly valuable for datasets like BreastMNIST, which aim to distinguish between benign and malignant tumors in breast tissue samples. Unlike other MedMNIST subsets, which often involve multi-class classification tasks, BreastMNIST focuses solely on distinguishing between benign and malignant tumors in breast tissue samples. This simplicity makes it an excellent starting point for anyone venturing into medical image processing and deep learning.

In this blog, we will explore why BreastMNIST is an optimal choice for beginners in medical AI research, walk through the initial steps of dataset handling, and discuss the potential improvements in AI agent integration for processing similar datasets in the future.

Why Choose BreastMNIST?

1. Binary Classification: With only two classes (benign or malignant), BreastMNIST reduces the complexity of training models, allowing researchers to focus on improving model performance without the added challenges of multi-class data.
2. Compact Dataset: It is one of the smallest datasets in the MedMNIST suite, making it easy to download, preprocess, and visualize without requiring extensive computational resources.
3. High Performance Potential: As shown in the comparison table below, BreastMNIST achieves impressive results across various architectures. Even with relatively fewer pixels, the dataset’s performance is competitive, highlighting its usefulness for algorithm testing and rapid prototyping.

As seen in the table, ResNet-18 on 28x28 images achieves the highest accuracy among the manually trained models, while Google AutoML Vision surpasses all other methods. Notably, increasing image size to 224x224 did not improve performance significantly, challenging the conventional expectation that higher resolutions yield better results. This finding suggests that the additional pixels may not provide meaningful new information for this specific task, and it raises broader questions about the diminishing returns of increased resolution in small datasets. For BreastMNIST, the task’s simplicity and the dataset’s design likely optimize performance at lower resolutions, emphasizing the need to carefully match model input sizes with the inherent characteristics of the data. These insights could influence future dataset curation and model development strategies.

Workflow: Processing and Training BreastMNIST

1. Dataset Download and Exploration: The dataset is freely available on the MedMNIST website or through Zenodo. No registration is required, ensuring easy access.
2. Visualization: Start by visualizing the 28x28 grayscale images to understand the data distribution. A simple Python script using libraries like Matplotlib can display these images. This step helps ensure that the data is correctly loaded and provides insights into class imbalances, if any. Once visualized, the dataset can be further reviewed for artifacts or inconsistencies.
3. Model Training: Using a lightweight convolutional neural network (CNN) such as ResNet-18 or a similar architecture, you can quickly achieve high performance. These models, optimized for simplicity and computational efficiency, are well-suited for BreastMNIST's binary classification task, enabling robust predictions with minimal resource requirements. Training metrics such as accuracy, AUC, loss curves, and confusion matrices should be visualized to evaluate the model’s robustness. Additionally, exporting these metrics to a JSON file can streamline analysis and facilitate sharing results.
4. Interactive Results: An interactive blog format would allow readers to scroll through sample images and view training metrics dynamically. By implementing scrollable image galleries, users can seamlessly explore multiple samples from the dataset without interruptions. Visualizations such as heatmaps for feature activation or plots comparing training and validation losses can be embedded directly, providing comprehensive insights into the model’s behavior. Additionally, interactive tools could allow users to hover over specific data points to view detailed metrics, making the analysis more engaging and precise. Furthermore, including animation transitions between visualizations might enhance clarity and maintain user engagement. This format could also support filtering options, enabling users to focus on specific metrics or subsets of data for tailored insights.

Explaining the Image Size Dilemma

Why does increasing image resolution from 28x28 to 224x224 fail to significantly boost performance? There are several potential reasons:

• Redundant Features: The additional pixels in higher resolutions may not provide meaningful new information for the task.
• Overfitting Risk: Larger input sizes increase the parameter count, making the model prone to overfitting, especially with limited training data.
• Preprocessing Alignment: The original MedMNIST dataset was curated with 28x28 images. MedMNIST+ now provides larger versions of the dataset, including resolutions of 64x64, 128x128, and 224x224 for 2D images, as well as 64x64x64 for 3D images. These larger sizes serve as standardized benchmarks for medical foundation models, which are generalized deep learning architectures pre-trained on large and diverse medical datasets. These models can be fine-tuned for specific tasks, making them valuable for advancing the performance of domain-specific applications like BreastMNIST. The original design of BreastMNIST with 28x28 images, however, remains advantageous for computational efficiency and simplicity in binary classification tasks.

Novel Training Methods for BreastMNIST

To further enhance the performance of models trained on BreastMNIST, novel training methods could be explored. These include:

• Self-Supervised Learning: Utilizing unlabeled data alongside BreastMNIST could allow models to learn generalizable features before fine-tuning on labeled data. Techniques such as contrastive learning or masked autoencoders could be applied to extract meaningful representations.
• Domain-Specific Transfer Learning: Instead of generic pretrained models, leveraging models fine-tuned on similar medical imaging tasks may improve performance. For example, initializing with weights from a model trained on mammography datasets could provide a head start for BreastMNIST.
• Curriculum Learning: Introducing examples in an order of increasing complexity might help the model learn better. Simpler, high-confidence examples could be presented first, gradually moving to more challenging cases.
• Data Augmentation Strategies: Beyond conventional techniques like rotation and flipping, advanced methods like GAN-based augmentation or mixup could generate realistic synthetic samples to increase data diversity.
• Explainable AI (XAI) Integration: Training methods that prioritize interpretability, such as attention-based mechanisms or prototype learning, could improve both performance and trustworthiness in medical AI applications.

By exploring these approaches, we could establish more robust and efficient training pipelines for BreastMNIST. Among these methods, self-supervised learning and domain-specific transfer learning are particularly impactful for immediate implementation. Self-supervised learning leverages unlabeled data to enhance feature extraction, while transfer learning from mammography-trained models accelerates performance improvements. These strategies offer a balance of practicality and innovation, making them well-suited for advancing the state of medical imaging research and application.

Conclusion

BreastMNIST stands out as an excellent dataset for binary classification in medical imaging, offering simplicity and computational efficiency without compromising on performance. Through innovative training methods, thoughtful exploration of dataset characteristics, and advancements in model architecture, researchers can unlock its full potential. The insights gained from working with BreastMNIST can guide broader medical AI applications, paving the way for enhanced diagnostic tools and improved healthcare outcomes. We encourage researchers and developers to explore this dataset and contribute to the growing field of medical imaging research.