机器学习开发代写|Data Mining 2023 Assignment 2: Prediction of Cellular Composition

这是一篇关于建立一个回归模型,预测给定组织图像补丁中六种不同类型细胞的数量以及图像补丁中细胞总数机器学习开发代写。具体作业要求是开发一个机器学习模型,该模型使用训练数据(具有给定细胞计数的补丁图像)来预测测试图像中每种类型的细胞计数以及每张图像中的细胞总数。在一个给定的图像补丁中不同类型细胞的计数称为其细胞组成。

 

In this assignment, the objective is to develop a regression model for predicting the number of six different types of cells in a given histological image patch as well as the total number of cells in the image patch. Your task is to develop a machine learning model that uses training data (patch images with given cell counts) to predict cell counts of each type in test images as well as the overall number of cells in each image. Counts of different types of cells in a given image patch is called its cellular composition.

We shall be using the data for the CoNIC: Colon Nuclei Identification and Counting Challenge (see:https://github.com/TissueImageAnalytics/CoNIC and https://arxiv.org/abs/2111.14485 for details about the challenge). However, you are not required or expected to participate in the formal challenge. The data files required for this coursework (counts.csv, images.npy) can be downloaded from: http://shorturl.at/fsGNU. This link also contains other files including a readme and binding licensing information for this dataset. You will also need to download the fold split file (split.txt) from

https://github.com/foxtrotmike/CS909/blob/master/2022/A2/split.txt.

You can read the data as follows:

import numpy as np

import pandas as pd

X = np.load(“images.npy”)#read images

Y = pd.read_csv(‘counts.csv’)#read cell counts

F = np.loadtxt(‘split.txt’)#read fold information

The data consists of 3 folds (in variable F). Here, X is a matrix containing N 256x256x3

Red-Green-Blue (RGB) channel images of size 256×256 pixels and Y is an Nx6 matrix with each row corresponding to a single image patch and each column corresponding to the 6 cell types (called T1: neutrophil , T2: epithelial T3: lymphocyte, T4: plasma , T5: eosinophil and T6: connective). Thus, for a single image, you are given the counts of each of the six cell types. The overall number of cells in an image can be obtained by summing up the number of cells of each type.

Training and Testing: Use data from the first two folds for training and validation and the third fold for testing. Wherever applicable, performance metrics for the test fold (3 rd Fold) are to be reported unless otherwise specified.

Submission: You are expected to submit a single Python Notebook containing all answers and code.

Include all prediction metrics in a presentable form within your notebook and include the output of the execution of all cells in the notebook as well so that the markers can verify your output. Also submit a consolidated table of your performance metrics within the notebook to indicate which model performs the best (MANDATORY).

Question No. 1: (Data Analysis) [20 Marks]

Load the training and test data files and answer the following questions:

i.How many examples are there in each fold? [2 marks]

ii.Show some image examples using plt.imshow. Describe your observations on what you see in the images and how it correlates with the cell counts of different types of cells and the overall number of cells. [2 marks]

iii.For each fold, plot the histogram of counts of each cell type separately as well as the total number of cells (7 plots in total). How many images have counts within each of the following bins? [4 marks]

0

1-5

6-10

11-20

21-20

31-40

41-50

51-60

61-70

71-80

81-90

91-100

>100

  1. Pre-processing: Convert and show a few images from RGB space to HED space and show the H-channel which should indicate cellular nuclei. For this purpose, you can use the color

separation notebook available here:

https://scikit-image.org/docs/dev/auto_examples/color_exposure/plot_ihc_color_separation.html [5 marks]

  1. Do a scatter plot of the average of the H-channel for each image vs. its cell count of a certain type and the total number of cells for images in Fold-1 (7 plots in total). Do you think this feature would be useful in your regression model? Explain your reasoning. [4 marks]
  1. What performance metrics can you use for this problem? Which one will be the best performance metric for this problem? Please give your reasoning. [3 marks]

Question No. 2: (Feature Extraction and Classical Regression) [40 Marks]

For the following questions, use the total number of cells as the output prediction variable.

  1. Extract features from a given image. Specifically, calculate the:
  1. average of the “H”, red, green and blue channels
  2. variance of the “H”, red, green and blue channels
  3. Any other features that you think can be useful for this work. Describe your reasoning for using these features.

HINT/Suggestion: You may want to use PCA Coefficients of image data (you may want to use randomized PCA or incremental PCA, see:

https://scikit-learn.org/stable/modules/generated/sklearn.decomposition.PCA.html). In case of computational complexity, you can reduce the number of images being used in determining the PCA basis. You can also look at other features such as GLCM features(https://scikit-image.org/docs/dev/auto_examples/features_detection/plot_glcm.ht

  1. ml) or transfer learning features. You can resize the images if needed. Plot the scatter plot and calculate the correlation coefficient of each feature in Q(2i,a-c) you obtain vs. the target variable (cell count) across all images. Which features do you think are important? Give your reasoning. [20 marks]
  1. Try the following regression models with the features used in part-I. Plot the scatter plot between true and predicted counts for each type of regression model for the test data.

Also, report your prediction performance in terms of RMSE, Pearson Correlation Coefficient, Spearman Correlation Coefficient and R2 score

(https://scikit-learn.org/stable/modules/classes.html#module-sklearn.metrics) on the test data. [20 Marks]

  1. Ordinary Least Squares (OLS) regression
  2. Support Vector Regression OR Multilayer Perceptron (MLP) OR Both

Question No. 3 (Using Convolutional Neural Networks) [40 Marks]

  1. Use a convolutional neural network (in PyTorch) to solve this problem in much the same was as in part (ii) of Question (2). You are to develop an architecture of the neural network that takes an image directly as input and produces a count as the output corresponding to the total number of cells. You are free to choose any network structure as long as you can show that it gives good performance. Report your results on the test examples by plotting the scatter plot between true and predicted counts on the test data. Also, report your results in terms of RMSE, Pearson Correlation Coefficient, Spearman Correlation Coefficient and R2 score. You will be evaluated on the design of your machine learning model and final performance metrics. Try to get the best test performance you can. Please include convergence plots in your submission showing how does loss change over training epochs.

[20 Marks]

  1. Use a convolutional neural network (in Pytorch) to predict the counts of 6 types of cells simultaneously given the image patch as input as well as the total number of cells (7 outputs in total). You are free to choose any network structure as long as you can show that it gives good cross-validation performance. Report the results for the test fold for each cell type in the form of separate predicted-vs-actual count scatter plots (3 folds, 6 cell types and 1 as the total number of cells so 21 plots in total) using your optimal machine learning model and report your results in terms of RMSE, Pearson Correlation Coefficient, Spearman Correlation Coefficient and R2 score for each cell type and the total number of cells. [20 Marks]

HINT:

A naïve (but possibly effective) strategy can be to simply try the same network architecture in 3(a) and train 7 different models separately for each cell type.

Restrictions:

Please do not share these data files or the assignment solutions publicly. Each student needs to submit a single solution and a solution should be developed by the student without assistance from other students. Students must adhere to the licensing information for this dataset.

Useful hints:

Feel free to resize the images to reduce the amount of required compute. However, if you do this,please ensure that the code for doing this is included in your submission notebook.

Look at glob (https://docs.python.org/3/library/glob.html) to get list of all file names in a given folder.

Look at skimage for image reading and image resizing (https://scikit-image.org/) and converting images to HED Space (see code below)

from skimage.color import rgb2hed

import skimage

from skimage.io import imread

print(‘skimage version’,skimage.__version__)

import matplotlib.pyplot as plt

I = X[0]/255.0 #read sample image and rescale pixel range in it

I_hed =

rgb2hed(I) #convert to HED

plt.imshow(I);plt.title(‘Original Image’);plt.show()

I_h = I_hed[:,:,0]; plt.figure(); plt.imshow(I_h,cmap=‘gray’);plt.colorbar();plt.title(‘H

Channel’);plt.show()

I_e = I_hed[:,:,1]; plt.figure(); plt.imshow(I_e,cmap=‘gray’);plt.colorbar();plt.title(‘E

Channel’);plt.show()

I_d = I_hed[:,:,2]; plt.figure(); plt.imshow(I_d,cmap=‘gray’);plt.colorbar();plt.title(‘D

Channel’);plt.show()

For calculating various regression metrics, please see:

https://scikit-learn.org/stable/modules/model_evaluation.html#regression-metrics


程序代写代做C/C++/JAVA/安卓/PYTHON/留学生/PHP/APP开发/MATLAB


本网站支持淘宝 支付宝 微信支付  paypal等等交易。如果不放心可以用淘宝交易!

E-mail: itcsdx@outlook.com  微信:itcsdx


如果您使用手机请先保存二维码,微信识别。如果用电脑,直接掏出手机果断扫描。

发表评论