Deep Learning-Based Positive Region Segmentation and Spatial Registration of Virtual Multiplex Immunohistochemical...

This study aims to facilitate virtual multiplexing by establishing a deep learning-based analytical framework for precise identification and spatial reconstruction of positive regions in multiple immunohistochemistry (IHC) images.

Twenty-two hematoxylin and eosin-stained whole-slide images (WSIs) and 86 serial single-marker IHC WSIs for virtual multiplexing from 80 mouse tumors, 50 Ki-67-stained images from the Multiplex IHC histopathological image classification database, and 100 from the Senaras2018DeepFocus data set were standardized using ImageJ software. Through image cropping and positive region annotation, an annotated data set for 4 indices (Ki-67, terminal deoxynucleotidyl transferase-mediated dUTP nick end labeling [TUNEL], tumor necrosis factor-α [TNF-α], intercellular adhesion molecule 1 [ICAM-1]) at 3 resolutions (512 × 512, 256 × 256, 128 × 128) was created. A pyramid-based multiscale mechanism was integrated into the U-Net architecture (Pyramid U-Net) to achieve efficient semantic segmentation of multiple IHC images. The SuperPoint feature extraction algorithm, SuperGlue matching algorithm, and random sample consensus algorithm were combined to achieve high-precision registration of virtual multiplex IHC. The DINO-ViTS16-WSI algorithm was integrated to map IHC-positive regions onto hematoxylin and eosin images for reconstruction and quantitative analysis.

The Pyramid U-Net model achieved high segmentation accuracy, maintaining a high mean intersection over union of 85.3% for Ki-67 and an accuracy of 97.9% for ICAM-1, while improving the Dice similarity coefficient for TUNEL by 5.8% to 85.4%. Computational efficiency increased by 5.14-fold, with single-image inference time for TNF-α reduced to just 24.3 ms. Validation based on both internal and external data sets demonstrated that the proposed model exhibited excellent segmentation accuracy and generalization ability. High-precision registration and spatial reconstruction were achieved, generating composite maps that visually displayed the spatial distribution of Ki-67 (proliferation), TUNEL (apoptosis), ICAM-1 (invasion), and TNF-α (inflammation).

This study describes a deep learning framework enabling accurate segmentation, registration, and spatial reconstruction of multiple IHC markers, which provides crucial technical support for tumor microenvironment analysis and personalized therapy via comprehensive multiparameter spatial analysis.