JCM | Free Full-Text | Digital Pathology Applications for PD-L1 Scoring in Head and Neck Squamous Cell Carcinoma: A Challenging Series

In the report of this educational program, we focused on the evaluation of the expression of the predictive tumor response marker PD-L1 via the CPS score in a digital series of selected challenging cases of HNSCC. For each case, a board of expert pathologists evaluated an initial original staining performed with an LDT protocol and a second CDx staining. Subsequently, a discussion in a face-to-face meeting was carried out to discuss the variability linked to the methods. This study highlighted discrepancies both in terms of technical and interpretation factors. As per the technical part, both abnormal nuclear/cytoplasmic PD-L1 staining (18%) and intensity variability with dirty backgrounds (27%) were noted, suggesting a role for both pre-analytical and analytical phases. Many factors can influence the performances of the PD-L1 IHC, including cold ischemia time, fixation, thickness, and age of histological sections or FFPE samples [1]. The thickness of the histological sections must be 4–5 μm, and, if unstained, these must not be older than 2 months to avoid the possible loss of immunoreactivity [1]. In terms of stored material, it is worth remembering that samples older than 12 months should be avoided, as they may show reduction in PD-L1 expression [1,12]. Factors such as DAB droplets, solitary dotting and other spots, background staining, and edge artifacts could instead be connected to technical aspects of the method, causing possible interference with the interpretation [1]. Fixation, processing, incomplete removal of paraffin, and incomplete rinsing of reagents from the slide affect the presence of background staining; a comparison with the control can help perceive the levels of these artifacts [1]. Considering the LDT clones, three out of eight cases stained with 22C3 and two out of three cases stained with the clone SP263 were associated with a technical discordance or critical issues, compared to the CDx Dako 22C3 pharmDx assay. Despite the use of the same antibody, platform, and detection kit, different parameters may vary among laboratories using alternative protocols, for example, the type and duration of antigen retrieval, the dilution of the primary antibody, the incubation time, and the amplification [13]. In addition, the Ventana platform produces a staining with a more intense immunoreactivity but with a background tissue that often appears “burnt/crushed” and is not perfectly preserved, which often makes it difficult to distinguish the various cellular types; however, with the Dako platform, the immunoreactivity is apparently milder, but there is a gain in terms of the morphological definition of the various cell populations. When using an LDT, validation with a standard method remains crucial [1,14]. Studies using tissue microarrays (TMA) have demonstrated that Dako 22C3 pharmDx and Ventana SP263 assays present similar distributions of PD-L1 expression in the tissue sections, although the second assay may produce more false-positive results due to stronger and more widespread staining at the tumor cell level [9,15]. From the interpretative point of view, the presence of unspecific DAB droplets may represent a confounding factor, leading to an underestimation of the CPS, as compared to preparations with cleaner backgrounds (cases n° 7 and 9). Staining intensity can represent another source of CPS variability, with lower stainings being a potential cause of underestimation, due to the reduced perception by the human eye (case 11). Moreover, abnormal nuclear/cytoplasmic chromogen localization (cases n° 2 and 3) can have repercussions on the formulation of the PD-L1 CPS score as well, especially in cases fluctuating around relevant cutoffs of 1 or 20, hampering final assignments above/below these values. However, it is possible to note that in our case series, all the interpretative problems involved PD-L1 CPS values very close to the cutoffs relevant for clinical purposes (i.e., 1 and 20). Various results have been reported regarding the agreement between pathologists in the formulation of the PD-L1 CPS score in HNSCC cases. If it is associated with adequate training, the reproducibility of the score can be high, reaching an intraclass correlation coefficient (ICC) even ≥0.70 [9], ≥0.80 [16], or ≥0.90 and even using various different assays (22C3 pharmDx, SP263, and SP142) [15,17]. In terms of clinical impact, here, interpretation variability could have been relevant either for therapeutic (case n° 2, CPS score 1) or prognostic (case n° 7, CPS from 15 to 22) purposes, again stressing the potential impact of the different staining techniques around sensitive cutoffs (LDT/board vs. CDx/board k = 0.63). To address this evaluation heterogeneity, digital pathology can be a natural solution for the assessment of CPS scoring [18]. Here, we adopted the intrinsic capabilities of QuPath to perform subjective qualitative or semi-quantitative evaluation to establish eventual differences in terms of DAB distribution/intensity among the different PD-L1 preparations in order to understand technical/interpretative discrepancies observed by human eyes. The data revealed significant differences in both comparisons between the two cellular subpopulations only in two cases (n° 9 and 11), in which the board also noted discordances in staining at the level of visual evaluation. The dissociation between computational results and the technical/interpretative differences noted by pathologists in the remaining cases can be explained by multiple factors. In particular, human evaluation is affected by both visual and cognitive traps, which are also present in the expert pathologist, as they are not linked to experience but are intrinsic to the human being, but they are absent in the computational data derived from the image analysis, since they are, by their own nature, quantitative and objective [19,20]. Here, computational methods were used retrospectively to understand technical/interpretative heterogeneity, but in a routine clinical context, they can be fundamental for the quality assessment of the IHC preparations performed with different methods and can be applied directly by pathologists for diagnostic/predictive purposes. This may be a game changer in the context of PD-L1 evaluation, since in many countries, it is not necessary to use the FDA-approved CDx; however, it is important only to use a test that has been validated for the specific aim [21]. Furthermore, it can prove to be an essential method for the evaluation and management of histological preparations aimed at new tools connected to the world of digital pathology, including artificial intelligence (AI) algorithms, in which the quality of the analyzed data is greatly affected by pre-analytical and technical variables.