Vectra ver 3

As previously described^{7 (link)}, four-micrometer-thick tissue sections obtained from formalin-fixed paraffin-embedded (FFPE) blocks were stained via mFIHC with an Opal IHC kit (AKOYA Biosciences, CA, USA). One representative FFPE block was selected from the pre-NAC and post-NAC specimens per case by the pathologist. The antibodies, dilutions, and activation conditions used are listed in Table S1. Next, a whole slide was scanned using an automated imaging system (Vectra ver. 3.0, AKOYA Biosciences). The whole specimens were captured, with an average of 20 areas at × 200 magnification. We segmented tumor tissues into cancer cell nests and stromal resions, and identified each stained cell with specific phenotypes using image-analyzing software (InForm, AKOYA Biosciences). Before the final evaluation, manual training sessions for phenotype recognition were conducted, followed by automatic machine learning for the algorithm. Two researchers (Ikarashi D, and S.T.) independently evaluated the stained slides, found no significant difference in result. An analytic program (Spotfire, TIBCO software, CA, USA) counted the infiltrating immune cells with specific phenotypes per mm² in cancer cell nests (intratumor) plus stromal regions (stroma) (Supplemental Fig. 2).

Multiplexed fluorescent‐labeled images of three randomly selected fields (669 μm × 500 μm each) were captured with an automated imaging system (Vectra ver. 3.0; Akoya Biosciences). An image analyzing software program (InForm; Akoya Biosciences) was used to segment cancer tissue into cancer cell nests (intratumoral) and the framework (stromal) region, and to detect immune cells with CD8 and T‐bet expression. Training sessions for tissue segmentation and phenotype recognition were repeated until the algorithm reached the level of confidence recommended by the program supplier (at least 90% accuracy) before performing the final evaluation.