WaveM-CNN for Automatic Recognition of Sub-cellular Organelles

This paper proposes a novel deep learning architecture WaveM-CNN for efficient recognition of sub-cellular organelles in microscopic images. Essentially, multi-resolution analysis based on wavelet decomposition and convolution neural network (CNN) are combined in the architecture. In each wavelet transformed sub-space, discriminative features are extracted by convolution kernels to provide various pattern characteristics of the same organelle. The generated feature maps are concatenated and passed directly to the fully connected layers of the classifier. In order to reduce the computational time and improve performance on limited dataset, transfer learning method is adopted, with the utilization of compact MobileNet model. Experiments on two benchmark datasets CHO and 2D HeLa are conducted to evaluate the performance of the proposed model on fluorescence microscopic images of sub-cellular organelles. The classification accuracies of 98.4% and 96.1% are achieved on these two datasets respectively, which are significantly higher than both hand-crafted feature based methods and recent deep learning based models.