Abstract
Multiple myeloma is a cancer that forms in a plasma cell which is a white blood cell. Usually, in multiple myeloma, cancerous plasma cells accumulate in the bone marrow and affect the healthy blood cells in multiple myeloma. Numerous methods are undergoing to build computer-assisted diagnostic tools for cancer diagnosis using various Image Processing techniques. These tools need capturing such images, stain color normalization of those images, then Cell segmentation and classification to count malignant versus healthy cells. Three deep learning models were explored to segment Multiple Myeloma (MM) plasma cells in this work. For this task, an official dataset provided as a part of the SegPC2021 competition was used. Moreover, three different kinds of models with various approaches are implemented. U-net architecture is used in the first approach, one of the most used neural network architectures for segmentation tasks. Moreover, with this U-net various backbones are used to get the best results. Here, the backbone refers to the extra network, which is helpful for feature extraction. In addition, various backbones such as ResNet, Densenet, EfficientNet are available for used architectures. This first approach has been evaluated on the SegPC2021 grand challenge dataset. As a second approach, Mask R-CNN architecture is used to improve the results of the first approach. A two-step deep learning method is developed as a third method. This method is considered the best and final method as it produced the best results for this work. In the first step, a nucleus detection network extracts each instance of the nucleus. After the instance is extracted, a multi-scale representation is generated using a multi-scale function. The primary purpose of the multi-scale function is to capture different sizes and shapes of cytoplasm instances. In the end, generated scaled images are passed into the network developed for cytoplasm detection.