<p>Multiple CT screenings are required to assess the progression of lung lesions, and this process involves accurately and precisely aligning two CT images. This is a tedious and difficult task for clinicians due to a number of numerous factors such as scanning device inconsistencies and body postures during the CT scanning process. This study introduces a Multi-depth-wise separable and Multi-scale feature guided Multi-resolution Projection and Excitation Affine Registration Network, MF-PEAR-net, for unsupervised 3D affine registration of medical images by optimizing the similarity loss. A multi-depth-wise separable kernel module is designed and used in place of the traditional convolution blocks to extract more extensive image features. A hierarchical multi-scale feature aggregation module is proposed fuse the extracted multi-scale features. The fused feature map is used to guide the re-calibration of the extracted features using a multi-scale feature guided project-and-excite module. Furthermore, the re-calibrated features at multiple scales are utilized by performing a dot product between two feature vectors, in order to improve the robustness of image representation. Extensive comparison experiments are performed on the proposed model and the recent state-of-the-art affine registration methods on two publicly available datasets that are pertinent to lung CT image registration, DIR-Lab and Learn2Reg. Results from quantitative and qualitative comparisons show that our model performs better than the single-step affine registration networks that are currently in use. The source code is available at: <a href="https://github.com/bonaldo112/MF-PEAR-Net/">https://github.com/bonaldo112/MF-PEAR-Net/</a>.</p>