The impact of site-specific digital histology signatures on deep learning model accuracy and bias

By Frederick M. Howard, J. Dolezal, S. Kochanny, Jefree J. Schulte, Heather Chen, L. Heij, D. Huo, R. Nanda, O. Olopade, Jakob Nikolas Kather and 3 more. Nature Communications, 2021

295 citations recorded by the source corpus.

The Cancer Genome Atlas (TCGA) is one of the largest biorepositories of digital histology. Deep learning (DL) models have been trained on TCGA to predict numerous features directly from histology, including survival, gene expression patterns, and driver mutations. However, we demonstrate that these features vary substantially across tissue submitting sites in TCGA for over 3,000 patients with six cancer subtypes. Additionally, we show that histologic image differences between submitting sites can easily be identified with DL. Site detection remains possible despite commonly used color normalization and augmentation methods, and we quantify the image characteristics constituting this site-specific digital histology signature. We demonstrate that these site-specific signatures lead to biased accuracy for prediction of features including survival, genomic mutations, and tumor stage. Furthermore, ethnicity can also be inferred from site-specific signatures, which must be accounted for to ensure equitable application of DL. These site-specific signatures can lead to overoptimistic estimates of model performance, and we propose a quadratic programming method that abrogates this bias by ensuring models are not trained and validated on samples from the same site. Deep learning models have been trained on The Cancer Genome Atlas to predict numerous features directly from histology, including survival, gene expression patterns, and driver mutations. Here, the authors demonstrate that site-specific histologic signatures can lead to biased estimates of accuracy for such models, and propose a method to minimize such bias.

Subfields: Artificial Intelligence, Oncology, Radiology, Nuclear Medicine and Imaging

View Paper: https://www.semanticscholar.org/paper/b399cc8e8d7a62a4d085b8893a153618ecefcc5f

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Oncology

In Radiology, Nuclear Medicine and Imaging

Citations in the corpus

Artificial intelligence in histopathology: enhancing cancer research and clinical oncology (2022, background): “Such differences include technical, digital and compositional biases of the training data 53 or even minor artifacts such as dust, scratches or fingerprints on slides 128 .”
Artificial intelligence for digital and computational pathology (2023, background): “Only through the concerted effort of multi-institutional data collection initiatives, open-source software packages and continued technical inspiration from advancements in computer vision and AI research can the said goals be reached.”
Algorithm fairness in artificial intelligence for medicine and healthcare (2023, background): “For example, medical images (such as radiographs, pathology slides and fundus photographs) can leak protected attributes, which can become ‘shortcuts’ to model predictions 9,12,89,90 .”
Comparing scientific abstracts generated by ChatGPT to real abstracts with detectors and blinded human reviewers (2023, unknown): “However, AI models have been shown to be highly sensitive to biases in training data 24,25 , and further data is needed to determine the potential for bias perpetuated by ChatGPT — especially given the overt prejudices emerging from prior language generation models 26 .”
Artificial intelligence in drug development (2025, unknown): “Applying AI and big data tools can effectively address pharmacometric problems and offer a powerful tool for time-to-event analysis, particularly in handling high-dimensional data and nonlinear relationships in the hazard function.”