Related Experiment Video
Updated: Jun 18, 2026
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
Methodological considerations for evaluating deep learning segmentation models in digital pathology whole-slide
Arian Arab1, Victor Garcia1, Seyed Kahaki1
1United States Food and Drug Administration, Center for Devices and Radiological Health, Office of Science and Engineering Laboratories, Division of Imaging, Diagnostics, and Software Reliability, Silver Spring, Maryland, United States.
Deep learning (DL) models for whole-slide image (WSI) segmentation in breast cancer show improved performance with color normalization, especially when data sources differ. Careful consideration of analysis units is crucial for accurate performance evaluation.
Area of Science:
- Computational pathology
- Digital pathology
- Artificial intelligence in medicine
Background:
- Automated whole-slide image (WSI) analysis using deep learning (DL) facilitates disease detection, classification, segmentation, and prognosis.
- Performance evaluation is critical for the success of big data technologies in pathology.
Purpose of the Study:
- To evaluate DL segmentation models on a breast cancer WSI dataset.
- To investigate methodological challenges in assessing WSI segmentation models.
- To assess the impact of color normalization on model performance across different data sources.
Main Methods:
- Evaluated DL model performance for segmenting tumoral and stromal regions in breast cancer WSIs.
- Assessed the effect of color normalization on model performance with cross-source data.
- Introduced three methods for aggregating segmentation performance (pixels, ROIs, slides).
- Employed a bootstrap method for estimating slide-level performance variance.
Main Results:
- Different units of analysis yielded varying mean performance estimates and uncertainty levels.
- Color normalization significantly enhanced DL model performance when training and testing data originated from different sources.
- The choice of aggregation unit impacts performance evaluation outcomes.
Conclusions:
- The study highlights the critical role of image acquisition, study design, and statistical analysis in evaluating computational pathology applications.
- Robust performance evaluation methodologies are essential for reliable DL model deployment in digital pathology.
- Standardized evaluation metrics and methods are needed for consistent assessment of WSI analysis tools.