As a popular deep learning model, the convolutional neural network (CNN) has produced promising results in analyzing lung nodules and tumors in low-dose CT images. However, this approach still suffers from the lack of labeled data, which is a major challenge for further improvement in the screening and diagnostic performance of CNN. Accurate localization and characterization of nodules provides crucial pathological clues, especially relevant size, attenuation, shape, margins, and growth or stability of lesions, with which the sensitivity and specificity of detection and classification can be increased. To address this challenge, in this paper we develop a soft activation mapping (SAM) to enable fine-grained lesion analysis with a CNN so that it can access rich radiomics features. By combining high-level convolutional features with SAM, we further propose a high-level feature enhancement scheme to localize lesions precisely from multiple CT slices, which helps alleviate overfitting without any additional data augmentation. Experiments on the LIDC-IDRI benchmark dataset indicate that our proposed approach achieves a state-of-the-art predictive performance, reducing the false positive rate. Moreover, the SAM method focuses on irregular margins which are often linked to malignancy.