Research: Batch Effects

Deep feature batch correction using ComBat for machine learning applications in computational pathology

Ever run experiments in different batches only to find mysterious variations in your results? In computational pathology, this common challenge becomes especially critical when AI models might learn lab-specific artifacts instead of actual biological signals.

Pierre Murchan et al. developed a solution by adapting ComBat harmonization - originally created for genomics data - to address batch effects in deep learning features from whole slide histopathology images.

The paper demonstrates how ComBat effectively reduces batch effects in AI models by removing confounding site-specific variations while preserving true biological signals.

When applied to colon and stomach adenocarcinoma datasets from The Cancer Genome Atlas, the method significantly decreased the model's ability to predict tissue source site (an indicator of batch effects) while maintaining strong performance on clinically relevant genetic features.

What makes this approach particularly valuable is its simplicity and effectiveness. By harmonizing deep features extracted from images, researchers can ensure their AI models learn actual disease patterns rather than artifacts from specific labs or imaging equipment. This addresses a critical challenge in deploying reliable AI tools for clinical applications.

As computational pathology continues advancing toward clinical implementation, addressing these hidden biases becomes essential for building trustworthy AI systems that generalize across diverse healthcare settings.

Research: Multi-modal Satellite

Using multiple input modalities can improve data-efficiency for ML with satellite imagery

Interesting research from Arjun Rao and Esther Rolf shows that using multiple data types can significantly improve machine learning efficiency with satellite imagery, especially when labeled data is limited.

They demonstrated that while most geospatial ML models rely solely on optical data like multi-spectral satellite imagery, adding complementary geographic inputs (like elevation maps, OpenStreetMap data, or environmental measurements) offers measurable benefits:

- Models trained with just 1-5% of available labeled data showed substantial accuracy improvements when using multiple data modalities
- Performance improved most significantly in out-of-sample settings (different geographic regions than training data)
- Two effective approaches were tested: simply stacking additional data as image bands, and using auxiliary tokens with Vision Transformers

The results are particularly relevant for applications where labeled data is often scarce and spatial generalization is crucial. Rather than always seeking more labeled data, researchers might see better returns by incorporating readily available geographic context into existing models.

Insights: Scanner Variability

Navigating Scanner Variability in AI-Powered Mitotic Figure Detection

Q: We know that different scanners result in a covariate shift reducing the AI performance in mitotic figure detection. Should we fine-tune the model for each scanner or retrain one model for use with all scanners?

The challenge of maintaining AI model performance across different medical scanners is a hot topic in computational pathology. 🔬 Here’s a deeper dive into the “it depends” answer from my recent webinar, expanded with cutting-edge insights:

𝐎𝐩𝐭𝐢𝐨𝐧 1: 𝐔𝐧𝐢𝐯𝐞𝐫𝐬𝐚𝐥 𝐌𝐨𝐝𝐞𝐥 𝐟𝐨𝐫 𝐀𝐥𝐥 𝐒𝐜𝐚𝐧𝐧𝐞𝐫𝐬
When it works:
- Large, diverse datasets are available (e.g., combining data from multiple institutions and scanner types).
- Limited training data from each scanner makes fine-tuning for each impractical.
- Stain augmentation methods are used to further take advantage of color variations.
Pros:
- Simplified deployment with one model.
- Cost-effective long-term maintenance.

𝐎𝐩𝐭𝐢𝐨𝐧 2: 𝐒𝐜𝐚𝐧𝐧𝐞𝐫-𝐒𝐩𝐞𝐜𝐢𝐟𝐢𝐜 𝐅𝐢𝐧𝐞-𝐓𝐮𝐧𝐢𝐧𝐠
When it works:
- Training data for each scanner enables fine-tuning.
- High precision is critical for a specific set of scanners (e.g., in regulatory contexts).
Pros:
- Maximizes performance for targeted workflows.
- Mitigates covariate shifts more aggressively.

𝐎𝐩𝐭𝐢𝐨𝐧 3: 𝐀 𝐇𝐲𝐛𝐫𝐢𝐝 𝐀𝐩𝐩𝐫𝐨𝐚𝐜𝐡
Start with a universal base model, then fine-tune lightly for individual scanners.
When it works:
- Large, diverse datasets are available to train a universal model.
- Training data for each scanner enables fine-tuning – although less data for each scanner will be needed than for Option 2.
Pros:
- Optimal performance for a wider range of scanners

𝐅𝐢𝐧𝐚𝐥 𝐓𝐚𝐤𝐞:
No one-size-fits-all answer exists. Prioritize data diversity and domain adaptation techniques for broad applicability, but invest in scanner-specific tuning when precision or regulatory demands outweigh scalability.

_{Enjoy this newsletter? Here are more things you might find helpful:}

_{1 Hour Strategy Session -- What if you could talk to an expert quickly? Are you facing a specific machine learning challenge? Do you have a pressing question? Schedule a 1 Hour Strategy Session now. Ask me anything about whatever challenges you’re facing. I’ll give you no-nonsense advice that you can put into action immediately.
Schedule now}

Did someone forward this email to you, and you want to sign up for more? Subscribe to future emails
This email was sent to _t.e.s.t_@example.com. Want to change to a different address? Update subscription
Want to get off this list? Unsubscribe
My postal address: Pixel Scientia Labs, LLC, PO Box 98412, Raleigh, NC 27624, United States