Author(s: Kriz Moses Deep learning A case study of Nucleus Segmentation using imaging experiments with Deep CNNs (UNet++, HRNet). Photo by National Cancer Institute. Table of Contents. Abstract. Introduction. Problem Statement. Literature Review. Challenges of medical picture segmentation. Sliding Window Method. UNet Exploratory Data Analysis. Deep Learning Architectures. Results. Discussion. Post-Training Quantization. Future. References. Because it allows you to identify every cell within a sample, identifying the nuclei of cells is the first step in most analysis. This process can be automated to save time and allow for faster drug testing. Deep CNNs are proposed to automate nucleus detection from images under different conditions. Three networks were created using U-Net++ and HRNet. I then compared their performance with Mean IoU. The U-Net model performed the best, with a mean IoU score 0 .861.. Image segmentation refers to the process of dividing an image into smaller sections for analysis and understanding. In the medical field, for example, doctors may want to highlight certain regions in brain scans (which are difficult to see with the naked eye) in order help them diagnose. A self-driving vehicle could also use this technique to help you segment various objects. Here is an example of how to segment blood vessels from retinal images. Retinal image and its segmentation mask (https://drive.grand-challenge.org/) In general, there are two types of segmentation: Semantic and Instance. Different classes of objects can be segmented in Semantic Segmentation (e.g. Separating background from people. Instance Segmentation is a method that separates different classes (e.g. separating different people from each other in the same picture) https://learnopencv.com/human-pose-estimation-using-keypoint-rcnn-in-pytorch/ Real-World Problem Statement and Dataset used In this case study, I strive to solve a semantic-based medical segmentation problem. This dataset is from the Kaggle competition 2018 Data Science Bowl. This dataset has been used to benchmark U-Net++/DoubleUNet. This dataset contains images of nuclei against different backgrounds. It is necessary to separate the nuclei from their background. Here are a few examples of the dataset. https://www.kaggle.com/c/data-science-bowl-2018/ Real-World Importance: The task that I will try to solve is to automate nucleus detection. It can speed up the research process and help to find a cure for nearly every type of disease. This includes heart disease and lung cancer, as well as rare diseases. Because most 30 trillion human cells have a nucleus, which is the genetic code that program each cell, identifying the nuclei is the first step in most analysis. By identifying nuclei, researchers can identify every cell within a sample. The researcher then measures how the cells respond to different treatments to understand the biological process [2].. The ML Formation of the Problem Image segmentation may also be posed as multiclass classification problem. Each pixel must be given a class. The input and output images are identical. My input data point is the cell image, while the output point the segmented nuclei and mask. An expert in this field labels the output image. The output image must be a binary with the background being black and the nuclei white. Given an input image of shape HxWx3, the task now is to create an output binary image of the segmented nuclear nuclei (shape. To assign a binary number between 0 and 1 to each pixel in the output binary picture. Segmentation of Biomedical Images in the Real World is crucial for diagnosis purposes. In some cases, even a slight error could lead to false diagnoses that can put patient’s health at risk. In our instance, the use of segmented nuclei for general treatment would mean that small mistakes shouldn’t be so significant. Latency does not require that the segments of nuclei be created in milliseconds. In most cases, it is not necessary to take more than a few seconds or even minutes. Use of Metrics The Mean Intersection Over Union is the metric. The Mean Intersection over Union. https://en.wikipedia.org/wiki/Jaccard_index Its a very commonly used metric along with dice coefficient. A mean IOU is the intersection area between two objects divided by their area of union. The two objects in image segmentation are the real region and the predicted regions. This is why we need to consider a football as our segment. The first object represents the area of the football as seen in the actual image. The second object is the area of football as shown in the image. What does the “Mean” in “Mean IIoU” mean? In “Mean IoU”, “Mean” refers to the average of all classes that are being segmented. In other words, we must segment the backgrounds from all cats in order to complete a task. This task could be described as a binary classification task. The two classes are a background and a cat. The mean IoU is calculated as IoU (cat) + background + 2; similarly, multiclass classifications have Mean IoU = IoU(1) + IoU(2) ….. + N. This reduces the effects of data imbalance. You can find a detailed explanation on Mean IOU here. Literature Review: Medical image segmentation challenges. Image segmentation was itself a challenging task for deep learning techniques because the output of the network must also be an image. CNNs typically take an image and then output a number. We had to produce an image. The problem was made even more difficult by the fact that there were very few training images for medical image segmentation. The lack of data makes this task extremely difficult. This task has been solved by traditional approaches such as clustering-based and thresholding. Deep learning has seen a lot of progress in medical image segmentation over the past few years. The sliding window method Each pixel of the output image can be predicted separately. Each pixel in the output image is fed with the surrounding cropped images (e.g 64x 64)) For each image HxWx3, we have three outputs: a total number of HxW outputs, and an 64x 64×3 section of the input picture as input. Although this method proved to be very simple at the time, it had many disadvantages. 1) The method took a long time to run and was not very efficient due to the overlapping of patches. Max-pooling was required when larger patches were used to maintain the global structure.
The 2018 Data Science Bowl: Medical Image Segmentation
