Abstract / Summary
Lymphomas comprise a diverse group of heterogeneous cell malignancies, primarily diagnosed through morphological and histochemical analysis of tissue biopsy specimens. Accurate automated nuclear segmentation is essential for quantitative histopathological analysis and downstream computational pathology applications, particularly in morphological heterogeneous lymphoma tissues. In the current study, we used deep learning-based approaches, namely a pre-trained StarDist model, a custom-trained StarDist model, and DeepLabV3, for automated nuclear segmentation of heterogeneous lymphoma cells in histological sections and achieved high accuracy of nuclear segmentation. DeepLabV3 achieved the highest pixel-level segmentation accuracy when trained on large annotated datasets, whereas StarDist demonstrated greater data efficiency under limited training conditions. We next compared the accuracy of nuclear segmentation between conventional formalin-fixed paraffin-embedded (FFPE) specimens and recently proposed glyoxal-based processing of rapidly frozen samples (CRYO). Overall, segmentation performance on CRYO specimens was comparable to that on FFPE specimens, although the magnitude and direction of differences varied by models. Custom-trained StarDist achieved a modestly higher Dice coefficient on CRYO specimens, whereas DeepLabV3 showed only a small, non-significant difference, and the pre-trained StarDist model showed no significant difference between tissue preparations. DeepLabV3 provided superior pixel-level segmentation accuracy, whereas StarDist demonstrated more accurate boundary localization, instance-level segmentation, and greater robustness to reduced training data. We conclude that both StarDist and DeepLabV3 can provide high-precision nuclear segmentation across FFPE and CRYO lymphoma specimens, supporting the applicability of CRYO processing for computational pathology workflows.