BackgroundMelanoma patients have shown sarcoid-like lesions as an immune-related adverse event to anti-PD-1 immunotherapy. Proper discrimination is essential for accurate treatment decision. Aim of the study is to distinguish granulomatous disease (GD) from pulmonary metastases (MET) and enlarged intrapulmonary lymph nodes (LN) in melanoma patients treated with immune checkpoint blockade by using deep learning. MethodsRetrospective (2012-2018) analysis of 165 melanoma patients treated with anti-PD-1 immunotherapy. All patients underwent standardized imaging-based tumor response assessment with contrast enhanced computed tomography (CT) (follow up interval 8-10 weeks) during treatment. Only patients with intrapulmonary lesions were included (n=132, f=72, median age 63 (IQR 55-74ys). All lesions were clustered in 4 different classes: (1) MET, (2) LN, (3) GD, and (4) small (2-5mm) indeterminate calcified granulomas (GC). GD was histologically proven (guided biopsy or operative excision). All LNs and GCs were identified in CT examinations (1-6 years) prior to diagnosis. For all lesions, a labeled region of interest was extracted by an experienced radiologist (11 years). Automatic classification of intrapulmonary lesions was performed by deep learning (VGG-like model). ResultsIn total, 4409 lesions were labeled (2746 METs (62.3 %), 1143 LNs (25.9%), 328 GC (7.4%), 192 GD (4.4%)). The model predicted instances of these classes for GD, GC or LN with a significant area under the receiver operating curve (ROC) of 0.66, 0.70 and 0.68, respectively (p-value <0.001). The performance of the algorithm was independent of imaging protocol (e.g. contrast enhancement, slice thickness, and reconstructed imaging kernel (soft versus sharp)) with AUC 0.65 for all. ConclusionsDeep learning can help to discriminate between intrapulmonary granulomatous disease, indeterminate calcified granulomas and intrapulmonary lymph nodes in anti-PD-1 treated melanoma patients. Further model development is needed to overcome the imbalanced “real clinical scenario” data to classify granulomatous disease in melanoma patients treated with anti-PD-1 immune checkpoint inhibitor blockade for implementation in the clinical workflow. Legal entity responsible for the studyThe authors. FundingHas not received any funding. DisclosureAll authors have declared no conflicts of interest.