Automated scoring of tumor-infiltrating lymphocytes informs risk of death from thin melanoma: A nested case-case study

To the Editor: Thin primary cutaneous melanomas (thickness ≤1.0 mm) have an excellent prognosis, with 10-year melanoma-specific survival rates of over 94%.1Garbe C. Keim U. Amaral T. et al.Prognosis of patients with primary melanoma stage I and II according to American Joint Committee on Cancer version 8 validated in two independent cohorts: implications for adjuvant treatment.J Clin Oncol. 2022; 40: 3741-3749Crossref PubMed Scopus (21) Google Scholar However, they constitute the majority of new melanoma diagnoses, and due to their frequency account for approximately one-quarter of total melanoma deaths.2Whiteman D.C. Baade P.D. Olsen C.M. More people die from thin melanomas (⩽1 mm) than from thick melanomas (>4 mm) in Queensland, Australia.J Invest Dermatol. 2015; 135: 1190-1193Abstract Full Text Full Text PDF PubMed Scopus (136) Google Scholar While accurate identification of high-risk patients could facilitate heightened surveillance and potentially systemic therapy, standard prognostic factors such as ulceration, mitotic rate, and sentinel node biopsy are rarely positive in thin melanoma, restricting their utility in stratification.3Claeson M. Tan S.X. Lambie D. et al.The association between BRAF-V600E mutations and death from thin (≤1.00 mm) melanomas: a nested case–case study from Queensland, Australia.J Eur Acad Dermatol Venereol. 2023; 37: e1168-e1172Crossref PubMed Scopus (3) Google Scholar We investigated whether an automated tumor-infiltrating lymphocyte (TIL) classification algorithm could predict mortality risk in thin melanoma. This nested case-case study was developed within a retrospective 20-year state-wide cohort of 27,660 patients with newly diagnosed thin melanoma. We randomly matched 85 fatal cases to 85 non-fatal cases according to age, sex, Breslow depth strata, year of diagnosis, and follow-up duration, as previously described [Supplementary Table 1, available via Mendeley at https://data.mendeley.com/datasets/9jhm839ck5/1].3Claeson M. Tan S.X. Lambie D. et al.The association between BRAF-V600E mutations and death from thin (≤1.00 mm) melanomas: a nested case–case study from Queensland, Australia.J Eur Acad Dermatol Venereol. 2023; 37: e1168-e1172Crossref PubMed Scopus (3) Google Scholar We evaluated all diagnostic hematoxylin-eosin slides through a QuPath-based cell classification neural network (titled NN192).4Acs B. Ahmed F.S. Gupta S. et al.An open source automated tumor infiltrating lymphocyte algorithm for prognosis in melanoma.Nat Commun. 2019; 10: 5440Crossref PubMed Scopus (57) Google Scholar,5Aung T.N. Shafi S. Wilmott J.S. et al.Objective assessment of tumor infiltrating lymphocytes as a prognostic marker in melanoma using machine learning algorithms.EBioMedicine. 2022; 82104143Abstract Full Text Full Text PDF PubMed Scopus (8) Google Scholar Cell type counts were quantified for each slide, and an electronic tumor-infiltrating lymphocyte % score (eTIL%) was calculated as TILs ÷ (tumor cells + TILs) [Fig 1]. For comparison, TIL grade (absent/brisk/nonbrisk) was manually coded by an expert dermatopathologist (D.L.). Conditional logistic regression, with adjustment for anatomical location, ulceration, and mitotic rate, was employed to determine odds ratios (ORs) and CIs for melanoma-specific mortality. Thin melanomas in the lowest eTIL% quartile demonstrated higher disease-specific mortality than tumors in the remaining quartiles (OR: 4.17; 95% CI: 1.71-10.16; P = .002), including after adjustment (adjusted OR: 4.77; 95% CI: 1.73-13.17; P = .003). At lowest quartile thresholding, eTIL% identified fatal cases of thin melanoma with a sensitivity of 0.36 and specificity of 0.86. On manual pathologist review, tumors with absent TILs exhibited a weak trend toward greater disease-specific mortality vs tumors with brisk TILs (OR: 1.69; 95% CI: 0.68-4.24; P = .262) and present TILs (nonbrisk and brisk; OR: 1.58; 95% CI: 0.77-3.26; P = .213); these remained nonsignificant after adjustment [Table I].Table IAssociation between histopathologic factors and electronic tumor-infiltrating lymphocyte % score with death from thin melanoma in a matched case-case series of 170 patients with primary thin melanomaFactorNonfatal (n = 85)Fatal (n = 85)UnadjustedAdjusted∗Adjusted for anatomical location (head and neck vs other), ulceration (present vs absent), and mitotic rate (≤1/mm2 vs > 1/mm2).OR (95% CI)POR (95% CI)PeTIL%, n (%) High (top 75%)73 (86%)54 (64%)RefRef Low (bottom 25%)12 (14%)31 (36%)4.17 (1.71-10.16).0024.77 (1.73-13.17).003Pathologist TIL, n (%) Brisk19 (22%)16 (19%)RefRef Nonbrisk51 (60%)47 (55%)1.09 (0.51-2.34).8171.06 (0.46-2.44).884 Absent15 (18%)22 (26%)1.69 (0.68-4.24).2621.68 (0.61-4.58).315Pathologist TIL, n (%) Present70 (82%)63 (74%)RefRef Absent15 (18%)22 (26%)1.58 (0.77-3.26).2131.60 (0.72-3.55).247Location, n (%) Other74 (87%)67 (79%)Ref Head and neck11 (13%)18 (21%)1.64 (0.77-3.47).198Ulceration, n (%) Absent83 (98%)79 (93%)Ref Present2 (2%)6 (7%)3.00 (0.61-14.86).178Mitotic rate, n (%) ≤1/mm283 (98%)71 (84%)Ref >1/mm22 (2%)14 (16%)7.00 (1.59-30.80).01P values calculated using conditional logistic regression. Results remain constant after a sensitivity analysis with adjustment for continuous Breslow depth [Supplementary Table 2, available via Mendeley at https://data.mendeley.com/datasets/9jhm839ck5/1].eTIL, Electronic tumor-infiltrating lymphocyte (as defined using the NN192 algorithm); OR, odds ratio; TIL, tumor-infiltrating lymphocyte.∗ Adjusted for anatomical location (head and neck vs other), ulceration (present vs absent), and mitotic rate (≤1/mm2 vs > 1/mm2). Open table in a new tab P values calculated using conditional logistic regression. Results remain constant after a sensitivity analysis with adjustment for continuous Breslow depth [Supplementary Table 2, available via Mendeley at https://data.mendeley.com/datasets/9jhm839ck5/1]. eTIL, Electronic tumor-infiltrating lymphocyte (as defined using the NN192 algorithm); OR, odds ratio; TIL, tumor-infiltrating lymphocyte. Despite the moderate correlation between pathologist TIL grade and eTIL% in this cohort (Spearman's ρ: 0.54), only eTIL% displayed prognostic significance. We postulate 3 potential reasons for this: firstly, by quantifying individual cells, eTIL% achieves higher granularity than ternary pathologic grading; secondly, eTIL% normalizes TIL count against the number of tumor cells (ie, tumor density), which is distinct from manual grading's emphasis on coverage across the vertical growth phase (ie, tumor area) and tumor base (ie, tumor length); thirdly, eTIL% includes peritumoral lymphocytes, which manual grading excludes. Overall, automated TIL analysis through the NN192 algorithm identified an immune-cold subset of thin melanomas at diagnosis that associated with higher melanoma-specific mortality beyond existing prognostic factors. Limitations of this study include retrospective sampling, the inability of hematoxylin-eosin staining to capture specific lymphocyte subsets, and the lack of data on comorbidities such as immunosuppression. Pending further optimization and prospective validation, automated TIL classifiers could readily complement existing pathologic workflows, with the potential to rationalize clinical management of early-stage melanoma according to patients' risk of disease progression. Dr Soyer is a shareholder of MoleMap NZ Limited and e-derm consult GmbH and undertakes regular teledermatological reporting for both companies; is a medical consultant for Canfield Scientific Inc, MoleMap Australia Pty Ltd, Blaze Bioscience Inc; and a medical advisor for First Derm. Dr Rimm has served as an advisor for Astra Zeneca, Amgen, Cell Signaling Technology, Cepheid, Danaher, Daiichi Sankyo, Konica Minolta, Merck, NanoString, PAIGE.AI, Regeneron, and Sanofi. Drs Tan, Aung, Claeson, Acs, Zhou; Author Brown; and Drs Lambie, Baade, Pandeya, Smithers, Whiteman, and Khosrotehrani have no conflicts of interest to declare.