Nodules are the primary biomarker for diagnosing and stratifying lung cancer malignancy. From CT scans, it is possible to characterize these abnormal masses, but analysis remains expert-dependent with remarkable challenges due to the typical morphological and textural variability. Computer-aided disease methodologies have recently shown capability support to classify lung nodules, yet they remain limited due to the following of classical supervised and balanced schemes, which results unrealistic, considering that less than 10% of the nodules are malignant. This proposed method takes advantage of benign nodule redundancy and trains a deep self-supervised representation following a one-class learning scheme. Then, malignant nodules are classified as anomaly reconstructed samples. The self-supervised methodology facilitates the extraction of spatial, local, and intermediate attention features, capturing texture patterns and enhancing the embedding robustness. An anomaly classification was conducted with the LIDC-IDRI dataset, achieving a 93.20% of AUC, demonstrating competitiveness regarding the state-of-the-art.Clinical relevance-This study approaches self-supervised representations to achieve generalizable nodule characterization, a promising alternative to transfer in clinical scenarios.