Hidden Bias in AI Cancer Diagnosis: How Training Data Skews Results and What Researchers Are Doing About It

Bias in AI Models Used for Cancer Diagnosis

Artificial intelligence (AI) tools hold promise for advancing medical image processing, but bias within these models can threaten their reliability. Professor Shahryar Rahnamayan, Chair of Engineering at Brock University, highlights that failing to address bias in medical AI can erode clinical trust and slow the adoption of technologies that might save lives.

Rahnamayan was part of a research team investigating bias in the Cancer Genome Atlas (TCGA)—a U.S.-based digital repository widely used to train and validate machine learning models in oncology. Their findings, published in Scientific Reports, reveal critical issues related to how these AI models learn from cancer tissue images.

Types of Bias Affecting AI Models

• Sampling bias: Occurs when training data does not accurately represent the diversity of the target population.
• Batch effect bias: Results from inconsistencies in how samples are collected, processed, or digitized across different healthcare institutions.

These biases can cause AI models to make inaccurate or harmful predictions. For instance, variations in staining techniques or imaging equipment create subtle signatures in biopsy images that AI can detect but human experts cannot. This means models might learn to associate certain image features with specific hospitals rather than the cancer itself.

Why This Matters for AI Diagnostics

When AI models pick up on hospital-specific characteristics, their accuracy drops when applied to data from other medical centers. Rahnamayan explains this as analogous to a doctor who can diagnose cancer accurately in one hospital but struggles in another due to unfamiliar conditions.

Since most machine learning models are trained on data from a limited number of hospitals, ensuring that these models generalize well across thousands of institutions is essential. Unaddressed bias undermines confidence in AI diagnostics and could limit their clinical utility.

Research Approach and Findings

The team hypothesized that the digitization process itself contributes significantly to the bias embedded in AI models. To test this, they developed a mathematical framework that converted color images into greyscale, isolating the influence of color on model behavior.

They found that greyscale images contain less bias than color images, which improves the generalization of deep neural network models across datasets. This finding suggests that reducing color-related variability can help create more reliable AI diagnostic tools.

Future research from the team will focus on identifying and mitigating various biases in medical AI applications to enhance both accuracy and generalization.

For those interested in expanding their knowledge on AI and healthcare, exploring advanced courses can provide practical insights into developing unbiased and trustworthy AI models. Visit Complete AI Training's latest AI courses for more information.