The term 'highly multiplexed imaging' refers to an imaging technology that produces multi- channel images. It is a crucial approach in biological sciences for medical applications, especially in cancer research. Imaging Mass Cytometry (IMC) is one such technology that allows for the concurrent detection of over 40 proteins. IMC images acquire cells by recording various protein expressions; up to 60 proteins can be captured, each of which is displayed as an image channel, and they intuitively reveal insights into specific diseases. To analyze IMC images, a key initial step is to identify and segment the cells among all the channels, which may then be utilized to investigate cell-cell interactions subsequently. The procedure of manually identifying these cells is difficult, time-consuming, and error-prone. Unfortunately, no methods for automatically identifying these cells in IMC images or interpreting the intrinsic relationships of distinct channel images are available. An application of IMC images is to distinguish tumor grades which can potentially play an important role in the clinical diagnosis, treatment, and prognosis of breast cancer. However, the potential for automated IMC image classification, using deep learning algorithms, has been developed and evaluated. The primary objectives of this thesis are to develop a set of deep learning algorithms to automate the analysis of IMC images. The IMC analysis includes segmenting cells of breast cancer tissue samples from the METABRIC cohort [1] and classifying tumor grades of breast cancer tissue samples from University Hospital Basel and University Hospital Zurich cohort [2]. For the cell segmentation task, a new Embedded Channel Attention Network (ECAN) is proposed to automatically identify cells in IMC images while utilizing an interpretation technique to explain inner channel importance. This fully supervised end-to-end segmentation network uses 2D convolutional kernels to extract features independently for each channel in the encoding steps, then 3D up-convolutional kernels in the decoding steps to encode the extracted high-dimensional features. The final segmentation results reveal that using the ECAN framework outperforms current state-of-the-art segmentation algorithms. To explain the importance of each channel, we present using a deep representational analysis of the internal features retrieved by ECAN as the channel importance Ranking Criteria. The results of the experiments indicate that applying our proposed method for channel importance ranking is consistent with biologically defined channel importance ranking. For breast cancer tumor grade classification, we propose a Channel-Transformer Classification Network (CTCN) that can accurately classify IMC images. The CTCN network is composed of the transformer's self-attention mechanism and the principle of channel embedding. Experimental results showed that the network with ResNest50 as the baseline provides the best classification accuracy.