Impact of scanner variability on lymph node segmentation in computational pathology

Computer-aided diagnostics in histopathology are based on the digitization of glass slides. However, heterogeneity between the images generated by different slide scanners can unfavorably affect the performance of computational algorithms. Here, we evaluate the impact of scanner variability on lymph node segmentation due to its clinical importance in colorectal cancer diagnosis. 100 slides containing 276 lymph nodes were digitized using 4 different slide scanners, and 50 of the lymph nodes containing metastatic cancer cells. These 400 scans were subsequently annotated by 2 experienced pathologists to precisely label lymph node boundary. Three different segmentation methods were then applied and compared: Hematoxylin-channel-based thresholding (HCT), Hematoxylin-based active contours (HAC), and a convolution neural network (U-Net). Evaluation of U-Net trained from both a single scanner and an ensemble of all scanners was completed. Mosaic images based on representative tiles from a scanner were used as a reference image to normalize the new data from different test scanners to evaluate the performance of a pre-trained model. Fine-tuning was carried out by using weights of a model trained on one scanner to initialize model weights for other scanners. To evaluate the domain generalization, domain adversarial learning and stain mix-up augmentation were also implemented. Results show that fine-tuning and domain adversarial learning decreased the impact of scanner variability and greatly improved segmentation across scanners. Overall, U-Net with stain mix-up (Matthews correlation coefficient (MCC) = 0.87), domain adversarial learning (MCC = 0.86), and HAC (MCC = 0.87) were shown to outperform HCT (MCC = 0.81) for segmentation of lymph nodes when compared against the ground truth. The findings of this study should be considered for future algorithms applied in diagnostic routines.