AI is rapidly becoming a key technology in cancer diagnostics, especially in digital pathology. By learning from high-resolution images of tissue samples, AI models help doctors detect tumors earlier, classify cancer types, and make better-informed treatment decisions. But there is one crucial step that rarely gets the attention it deserves: cleaning the data.
Before an AI model can make accurate predictions, it needs to be trained on data that is clean, consistent, and trustworthy. That behind-the-scenes effort, including removing noise, correcting labels, and standardizing images, is what truly sets high-performing cancer AI apart.
Why Raw Pathology Data Isn’t Ready for AI (Yet)
Histopathology images, or whole slide images (WSIs), are incredibly detailed. They often exceed a billion pixels and capture subtle features in cancerous tissue. But they also come with a host of hidden issues:
-
Color inconsistencies from different staining methods and lab equipment [1]
-
Physical artifacts like tissue folds, pen marks, and blurry regions [2]
-
Incorrect or inconsistent labels that can throw off the AI’s learning process [3]
Even the best AI algorithms struggle if the data is full of distractions. Studies have shown that these issues can significantly reduce model accuracy and even introduce unintended bias [4].
How Data Cleaning Makes AI Smarter and Safer
To make cancer datasets AI-ready, teams of scientists and engineers use smart preprocessing techniques that help models focus on what matters:
-
Stain normalization adjusts color differences across slides, reducing confusion [1]
-
Artifact detection identifies and filters out unwanted image distortions [2]
-
Label verification ensures that diagnostic labels are consistent and reliable [3]
-
Outlier filtering removes data that doesn’t fit or is duplicated [4]
The impact of these steps is real. One study showed that correcting mislabeled slides increased a model’s performance by over 25% [5]. Another found that filtering artifacts significantly improved how well the AI worked across different hospital datasets [2].
Clean Data Builds Trust and Better Outcomes
AI in healthcare doesn’t just need to be smart. It needs to be trusted by doctors, regulators, and patients alike. That trust starts with high-quality, well-curated data.
Clean data ensures that models perform reliably, not just in a single lab, but in clinics and hospitals around the world. It reduces the risk of misdiagnoses, accelerates regulatory approval, and increases adoption by healthcare professionals who want to understand how these tools actually work.
From a business standpoint, it also saves time and resources. AI tools built on solid data foundations are more performant, less likely to fail in validation, need fewer revisions, and get to market faster.
Want Better Cancer AI? Start With Better Data
At the heart of every accurate, generalizable, and ethical AI tool is one thing: high-quality data. And while data cleaning may not make headlines, it’s what enables AI to learn from real pathology, not noise or artifacts.
If you are curious about what makes a truly reliable cancer dataset, which is not just clean, but also diverse, well-annotated, and scalable, we explored those bigger-picture factors in our recent piece: What Makes a Good Cancer Dataset for AI?
At PAICON, we believe that precision oncology starts with precise data. It is how we build AI that not only detects cancer but helps defeat it.
References
-
Artificial Intelligence in Clinical Oncology: From Data to Digital Pathology and Treatment. https://ascopubs.org/doi/10.1200/EDBK_390084
-
Unleashing the potential of AI for pathology: challenges and recommendations. https://pmc.ncbi.nlm.nih.gov/articles/PMC10952719/
-
Artificial intelligence in digital pathology: a systematic review and meta-analysis of diagnostic test accuracy. https://www.nature.com/articles/s41746-024-01106-8
-
Artificial intelligence in diagnostic pathology. https://diagnosticpathology.biomedcentral.com/articles/10.1186/s13000-023-01375-z
-
Preparing Data for Artificial Intelligence in Pathology with Clinical-Grade Performance. https://pmc.ncbi.nlm.nih.gov/articles/PMC10572440/
