P1000

Sixty-eight samples were obtained from patients who suffered from hepatocellular carcinoma and underwent hepatectomy in The First Affiliated Hospital, Zhejiang University School of Medicine. Paraffin-embedded tissue was sliced and dewaxed. After antigen retrieval, primary antibodies were incubated with slides at 4°C overnight. Herein, CD138 (rabbit monoclonal to Syndecan-1/CD138, Cat. No. ab128936, Abcam, United Kingdom), CD86 (rabbit polyclonal to CD86, Cat No. DF6332, Affinity Biosciences, USA), and PD-L1 (mouse monoclonal to PD-L1/CD274, Cat No. 66248-1-Ig, Proteintech, USA) were used. After washing, secondary antibodies were incubated at 37°C for 30 min and washed. Then diaminobenzidine (DAB) was applied for color development. Lastly, all sections were scanned by Pannoramic DESK, P-MIDI, P250, P1000 (3D HISTECH; Hungary) and were read by Pannoramic Scanner (3D HISTECH; Hungary).
All sections were evaluated by three pathologists independently. CD138 and CD86 were categorized into positive (+) or negative (−), and PD-L1 was scored ranging from 0 to 3 (0, 1, 2, and 3). If there was disagreement of the evaluation, the results of CD138 and CD86 adhered the principal of minority obeying majority, and the final score of PD-L1 was the average of values from three observers. The score of PD-L1 was categorized into two groups (high and low) according to the median value.

Skin specimen WSIs that met the inclusion criteria were stained as previously described, digitized using 1 of 3 models of 3D HISTECH digital slide scanners (P100, P250, or P1000; 3D HISTECH, Budapest, Hungary). As images are normalized per pixel and data augmentation applied to offset variation in color values, variation in scanning devices only serves to diversify the Training Set and strengthen the Validation Set. The resulting .mrxs files were annotated by human annotators using the open-source pathology and bioimage analysis software QuPath (v0.1.3.5)⁶⁴ (link) to create ground-truth labels for SL (See “Datasets: training set”), as well as in other downstream processes.

The tongues of the mice were harvested and cut into pieces after 1 week, and the tissues were classified according to the type of lesion. One part was fixed in 10% neutral-buffered formalin and embedded in paraffin for IHC analysis. The other part was frozen in an optimal cutting temperature embedding compound, and $30\text{-}\mu \mathrm{m}$ sections were cut for fluorescence imaging using Pannoramic DESK, P-MIDI, P250, and P1000 (3D HISTECH) at the 800-nm channel with $5\text{-}\mu \mathrm{m}$ resolution. For HE staining, serial $5\text{-}\mu \mathrm{m}$ sections were cut.

For $\alpha \mathrm{v}\beta 6$ and EGFR protein quantification, IHC slides were digitalized using a Pannoramic DESK, P-MIDI, P250, and P1000 (3DHISTECH, Budapest, Hungary). In the surgical specimens, we separately analyzed tumor and normal tissues. Thereafter, $\alpha \mathrm{v}\beta 6$ and EGFR were quantified using Image J/FIJI software for at least three fields of view per area of interest (tumor and normal tissue) ( $20\times$ magnification).
Diaminobenzidine and HE staining were separated using the color deconvolution algorithm, and appropriate threshold levels were set to measure the area of specific diaminobenzidine staining ( $\alpha \mathrm{v}\beta 6$ area) and HE staining (tissue area) to calculate the relative $\alpha \mathrm{v}\beta 6$ positive area on each image. The thresholds were kept constant within each dataset. We recorded the mean and standard deviation for each sample. For grouped analysis, we used a paired $t$ -test.