Impact of scanner variability on lymph node segmentation in computational pathology.

Area of Science:

• Digital Pathology
• Computational Histopathology
• Medical Image Analysis

Background:

• Digitization of glass slides is fundamental for computer-aided diagnostics in histopathology.
• Image heterogeneity from different slide scanners can impair computational algorithm performance.
• Accurate lymph node segmentation is critical for colorectal cancer diagnosis.

Purpose of the Study:

• To evaluate the impact of slide scanner variability on lymph node segmentation performance.
• To compare different segmentation methods under scanner variability conditions.
• To assess the effectiveness of domain generalization techniques in mitigating scanner effects.

Main Methods:

• Digitization of 100 lymph node slides using 4 different scanners.
• Annotation of 276 lymph nodes by experienced pathologists.
• Application and comparison of Hematoxylin-channel-based thresholding (HCT), Hematoxylin-based active contours (HAC), and U-Net (convolution neural network).
• Implementation of fine-tuning, domain adversarial learning, and stain mix-up augmentation for U-Net.
• Evaluation using Matthews correlation coefficient (MCC).

Main Results:

• Fine-tuning and domain adversarial learning reduced scanner variability impact, enhancing cross-scanner segmentation.
• U-Net with stain mix-up (MCC=0.87) and domain adversarial learning (MCC=0.86), along with HAC (MCC=0.87), outperformed HCT (MCC=0.81).
• These advanced methods demonstrated improved robustness against scanner-induced image variations.

Conclusions:

• Scanner variability poses a challenge for computer-aided histopathology diagnostics.
• Fine-tuning and domain adversarial learning are effective strategies to improve algorithm generalizability across different scanners.
• The findings support the integration of robust segmentation methods in diagnostic routines for improved accuracy and reliability.