Figure 1.

Architecture of MFCN. The MFCN includes three modules: a multi-scale inception module, a multi-branch sampling module and a multi-scale ensemble module. The multi-scale module uses three different kernel size convolutional layers (3 × 3, 5 × 5, 7 × 7, stride 2) to extract features, which are expanded to form branches concatenated with other modules, respectively. In multi-branch module, we have combined three branches (blue, green and light red) with different steps of down-sampling and up-sampling (blue: 16× down-sampling and up-sampling; green: 8× down-sampling and up-sampling; light red: 4× down-sampling and up-sampling). Finally, the ensemble module combines multi-mode features extracted from the multi-branch module to detect vesicles with different sizes and shapes. The number of channels of feature maps is denoted on top of the cuboids and the arrows denote the different operations (kernel size/stride).