The urine-formed element instance segmentation is of great importance for the diagnosis of urinary system diseases. However, current instance segmentation methods tend to lose detailed information for small cell targets. Moreover, mutual occlusion of the cells is more likely to result in error segmentation such as missed detection, false detection, and over-segmentation. To address the above problems, an improved instance segmentation method called MSCD R-CNN is proposed. MSCD R-CNN adopts the ResNet 50/101-CBAM-FPN as its backbone network. It incorporates the convolutional block attention module, which makes the network focus on the objects properly and improves detection and segmentation quality. Additionally, MSCD R-CNN designs a generative adversarial network head called MaskDis Head to improve segmentation performance by adversarial training between predicted and real masks. To assess MSCD R-CNN performance, we conducted experiments on the test dataset and adopted standard COCO evaluation metrics. The experimental results show that the MSCD R-CNN method has an AP@0.50 of 91.84%, an AP@0.50-0.95 of 55.55%, an AR@0.50 of 96.49%, and an AR@0.50-0.95 of 62.25% on the test dataset. Compared with Mask Scoring R-CNN (MS R-CNN), the AP@0.50 and AP@0.50-0.95 are improved by 2.94 percentage points and 2.30 percentage points, respectively, and the AR@0.50 and AR@0.50-0.95 are improved by 4.24 percentage points and 0.98 percentage points, respectively. The experimental results demonstrate that the MSCD R-CNN achieves state-of-the-art performance compared with other advanced mainstream methods. These results confirm the effectiveness of the MSCD R-CNN in accurately segmenting the formed elements in urine, providing good technical support for clinical disease diagnosis.