Reconstructing physical field through data-driven deep learning has gained increasing attention. However, it's quite time-consuming or costly on creating sufficient samples for a three-dimensional complex problem. To improve the accuracy of reconstructed physical field for a complex geometry with a small dataset, segmented-combined learning (SCL) framework is proposed for designing the prediction model in this paper. Instead of directly constructing the physical field of the complete geometry, the core idea of SCL is that a prediction model contains multiple networks, each of which learns to construct physical field of the assigned segment of geometry with local geometric images and operating parameters as its designed input. The learned physical fields of segments are then combined to realize the complete reconstruction. Herein, the SCL is described in detail and validated with the physical field reconstruction of laminar flow in a circular tube with random inserts. A basic conditional Generative Adversarial Networks (cGAN) model and an SCL-based cGAN model are established to reconstruct velocity magnitude and temperature fields, which are trained with the same hyperparameters and small dataset created by Computational Fluid Dynamics (CFD). To achieve better results, a few training strategies are adopted, and several hyperparameters are optimized. The test results show that compared with the basic cGAN model, Mean Squared Error (MSE) in pixel values of the reconstructed physical fields of the SCL-based model is reduced by 42.47%, to 81.64, and Structural Similarity (SSIM) is improved by 5.24%, to 89.24%, which proves that SCL is very effective. This study provides a new way for improving the accuracy of reconstructed physical field with a small dataset.