Multiple myeloma (MM) is a blood cancer that develops when plasma cells expand abnormally in the bone marrow. The early detection of MM is beneficial for accurate treatment in time and draws increasing recognition. There are several methods to detect myeloma cells in bone marrow, such as using microscopic analysis based on the aspirate slide images. In this paper, we propose a deep learning framework called the semantic cascade Mask RCNN for the detection and segmentation of myeloma cells. The framework is also integrated with the proposed feature selection pyramid network, which is a simple and effective module to improve the segmentation performance. The mask aggregation module refines and merges the high certainty instance masks into a single segmentation map and combines the results from the extra semantic segmentation branch to generate better predictions. The extensive experiments on the SegPC-2021 Challenge dataset demonstrate that the proposed method achieves a promising performance.