Abstarct: Computer-aided analysis of microscopic images plays an important role in the diagnosis and prognosis of diseases. Due to the large volume of image data, manual processing is very difficult and time consuming or due to fatigue is associated with errors. Therefore, increasing the speed of clinical diagnosis through microscopic images and automating the analysis of images effectively and with high accuracy is very important. The high resolution of the microscopic images, the large number of important objects in these images, the crowdedness of the image and the possibility of overlapping objects, the high similarity of some objects to each other, the appearance of each object in different shapes in different parts of the image, the existence of artifacts and data imbalance are the important challenges in this field. The aim of this thesis is to present a method baxsed on artificial intelligence and deep learning techniques for the analysis of histopathological images to diagnose tissue diseases, mextastasis, prognosis of diseases and to help determine the progress of diseases. In this thesis, automatic and accurate microscopic diagnosis of anthrax through the diagnosis and segmentation of Bacillus anthracis bacteria and important cells of the immune system, including lymphocytes, macrophages and neutrophils despite the challenges in this field, is investigated. According to studies, an efficient algorithm for automatic microscopic diagnosis of anthrax and segmentation of its microscopic images has not been presented yet. In order to improve the segmentation efficiency and solve the mentioned challenges, deep learning models baxsed on semantic segmentation including Attention-IRUNet (with higher segmentation accuracy), improved UNet++ (with higher segmentation speed) and URCNN which is a model baxsed on instance segmentation are presented. Besides, for each of the proposed models, an appropriate and effective loss function is provided. Finally, in order to use the strengths of these models simultaneously and fix their weaknesses, a hybrid model baxsed on the gradient boosting learning algorithm has been used to combine the results predicted by two important proposed architectures including Attention-IRUNet and URCNN. The Attention-IRUNet and improved UNet ++ models take advantage of the combination of feature maps of different scales to enhance the feature maps at each scale, as well as the squeeze and excitation-residual and squeeze and excitation-inception blocks to improve the representation quality of the features. In addition, in the Attention-IRUNet model, Attention Gate blocks are used to focus the model more on the target structures, and deep supervision technique is used to improve the learning ability of the models. In order to solve the problem of data imbalance of different classes and focus more on hard or incorrectly classified samples, a weighted hybrid loss function is also proposed. In the proposed URCNN model, which is an improvement of the Mask-RCNN deep architecture, a feature pyramid network is proposed by modifying the FPN, which improves detection and segmentation accuracy. Also, in order to improve the segmentation efficiency in this model, the use of a hybrid loss function and a U-shaped structure with multi-scale skip connections in the mask branch is suggested. The experimental results show that despite the challenges in the field of microscopic image analysis, the proposed models, especially the hybrid model have performed better (with values of Precision, Recall, IoU and Dice equal to 85.9%, 88.2%, 76% and 85.9%, respectively) in the segmentation of microscopic images compared to the superior models presented in the recent years papers.