This study aims to segment the maxillary sinus into maxillary bone, air, and lesion, by employing an active learning framework in conjunction with a customized 3D nnU-Net. Tested on 83 randomly selected cases, the framework uses limited cone-beam computed tomography (CBCT) datasets, reducing annotation efforts and costs. The approach shows consistent accuracy for air and lesion segmentation across three progressive stages, as indicated by the Dice similarity coefficients (DSCs). The use of this framework also significantly decreases the time required for segmentation. Overall, the findings demonstrate that the deep active learning framework can efficiently train on limited CBCT datasets, potentially enhancing segmentation tasks in dental CBCT images. This enhancement is particularly noticeable when using Coreline Soft's AVIEW Modeler software.