With the rapid development of fluorescence microscope technologies, high-content screening and light-sheet microscopy are producing ever-larger datasets that pose great challenges in data storing and data sharing. As a popular compression tool, B3D introduces a noise dependent compression algorithm for microscopy images to preserve the numerical intensities of all pixel within their uncertainties by exploiting the natural variability of each pixel value. Nevertheless, the high complexity of the processing flow restricts the deployment of the tool since the throughput and power consumption cannot satisfy the increasing demand. In this paper, we propose an efficient FPGA-accelerated data compression framework based on B3D. Following the co-design methodology, the compression processing flows are partitioned into different blocks to deploy on CPU or FPGA according to their computation characteristics. Also, we design a custom accelerator core that consists of multiple full on-chip pipelines using the channel function of the Intel OpenCL toolkit to implement data-flow driven computation. Our experiments show that the proposed framework achieves up to 32× throughput for a single pipeline compared with Intel Xeon E3-1220 v5 operating at 3.00 GHz, and 6× energy-efficiency compared with GPU implementation.