1214 AI-derived histopathological features predict survival outcomes in advanced melanoma
Year: 2024
Type: conference-abstract
Abstract: Background In the USA, melanoma ranks as the second and fourth most prevalent cancer among males and females respectively.1 2 Recent advances in immunotherapy have increased melanoma survival rates; however, about 50% of metastatic melanoma patients do not respond to this treatment.3 There is an urgent need to identify factors and tools that can identify patient subgroups that are likely to benefit from immunotherapy. We use artificial intelligence (AI) computer vision models to analyze the tumor microenvironment (TME) in hematoxylin and eosin (H&E) stained whole slide images (WSIs), extracting quantifiable features such as tumor micron area to predict overall survival (OS) and progression-free survival (PFS) in advanced melanoma. Methods We sourced Formalin-Fixed Paraffin-Embedded (FFPE) tissues from 152 melanoma patients archived at the Pathology Unit and the Melanoma Cancer Immunotherapy and Innovative Therapy Unit of the Istituto Nazionale Tumori IRCCS ‘Fondazione G. Pascale’, Naples (Italy). Each patient had a WSI with metadata including OS, PFS, treatment, and response. All WSIs passed a quality control (QC) assessment performed by our board-certified pathologist. A subset of 85 WSIs were manually annotated by our pathologist for the following regions: Tumor, Stroma, Necrosis, Normal, and Lymphocyte cluster. A machine learning model was developed using these annotations to predict heatmaps with these five tissue classes. Heatmaps for all 152 WSIs underwent QC checks by our pathologist, and model development iterated until all WSIs passed. For every WSI we extracted area features for each tissue class by summing the total area covered by the heatmap. We then computed the ratios between the areas of each pair of tissue classes. Patients were divided into ‘below median’ and ‘above median’ groups for each feature, with box plots visualizing the data distribution, Kaplan-Meier curves plotted for OS and PFS, and log-rank tests performed to assess significance. Results One feature describes the ratio between the area of ‘Normal’ tissue and ‘Lymphocyte cluster’ tissue; using a median value of 16.79, we observed a significant difference in OS (p=0.028) and PFS (p=0.008). Additionally, significant differences in OS (p=0.011) and PFS (p=0.023) were observed for the ratio between the area of ‘Stroma’ tissue and ‘Lymphocyte cluster’ tissue, with a median value of 1.77. Conclusions Our findings identify two AI-derived quantitative features from WSIs of advanced melanoma patients that are predictive of survival outcomes. These features hold promise as tools to enhance the selection and personalization of immunotherapy treatments. Acknowledgements The authors thank the patients and families who made this trial possible. In addition, the authors thank the Italian Ministry of Health (IT-MOH) for having partially supported the present study by ‘Ricerca Corrente’ (grant no. L2-1). References Miller KD, Nogueira L, Devasia T, Mariotto AB, Yabroff KR, Jemal A, Kramer J, Siegel RL. Cancer treatment and survivorship statistics. CA Cancer J Clin. 2022. Forsea AM. Melanoma epidemiology and early detection in Europe: diversity and disparities. Dermatol Pract Concept 2020;10:e2020033. Shalata W, Attal ZG, Solomon A, Shalata S, Abu Saleh O, Tourkey L, Abu Salamah F, Alatawneh I, Yakobson A. Melanoma management: exploring staging, prognosis, and treatment innovations. Int J Mol Sci. 2024 May 26;25(11):5794. Ethics Approval The study was approved by Istituto Pascale’s Ethical Committee with the protocol DSC 33/22 oss (all patients signed an informed consent, which allowed the use of clinical data for research purposes and analysed in an anonymous manner). The study design and conduct complied with all current regulations regarding the use of human study participants and was conducted in accordance with the criteria set by the Declaration of Helsinki.
Language: English
Topic: AI in cancer detection
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4404064952