AI Models Prone to Errors Due to Tissue Contamination, Finds Study


Artificial intelligence (AI) models, which are typically trained in controlled environments, are prone to errors when exposed to real-world scenarios involving tissue contamination, according to a recent study by Northwestern Medicine. The study highlights the challenges faced when lab-trained AI is confronted with the diversity of materials it has not been trained on. The findings serve as a reminder that even though AI performs well in the lab, it may struggle in real-world applications. Therefore, the final decisions on diagnoses made on biopsies and other tissue samples should ultimately be made by human experts.

In this study, researchers trained four AI models to analyze microscope slides containing placenta tissue. The models were trained to (1) detect blood vessel damage, (2) estimate gestational age, (3) classify macroscopic lesions, and (4) detect prostate cancer in needle biopsies. The scientists then exposed each model to small portions of contaminant tissue randomly sampled from other slides to gauge their reactions.

The study revealed that all four AI models paid excessive attention to tissue contamination, resulting in errors in diagnosing vessel damage, gestational age, lesions, and prostate cancer. This is the first study to examine the impact of tissue contamination on machine-learning models in the field of pathology. Tissue contamination is a well-known issue for pathologists, who are trained to ignore contaminants, but it often catches non-pathologist researchers and doctors by surprise.

When human pathologists examine tissue on slides, they focus on a limited field within the microscope. After examining the entire sample, they combine all the information gathered to make a diagnosis. However, AI models can be easily misled by contaminants. The models struggled to decide which elements to pay attention to, resulting in a distraction from the actual tissue being examined.

The study found that AI models gave significant attention to contaminants, indicating their inability to distinguish biological impurities. Researchers highlight the need for practitioners to quantify and improve upon this issue. While previous studies have investigated image artifacts like blurriness or debris on slides, this is the first study to examine the impact of tissue contamination.

The study concludes that AI models should not be considered as replacements for human expertise in pathology. Patients should continue to rely on human pathologists as the final decision-makers in diagnoses made on biopsies and tissue samples. Although AI has the potential to revolutionize the healthcare industry, it is crucial to address the limitations and challenges it faces in real-world scenarios. As technology continues to evolve, AI models can be refined to improve their accuracy and reliability, ensuring they augment the capabilities of healthcare professionals rather than substituting them.

1. Source: Coherent Market Insights, Public sources, Desk research
2. We have leveraged AI tools to mine information and compile it