Biomedical classification

Technology

• CellProfiler
• Python 3.10.4
• pandas==1.4.3
• seaborn
• torch==2.1.1 : Neural Network
• Scikit-learn: Logistic Regression, Multilayer Perceptron

Reason

• Obtaining better results than traditional CNN which achive only 62% accuracy because of:
  • difficult nature of the data
  • various data distributions
  • small number of individuals

Data

• The original data are microscopic images of the brains of mice aged 4, 8 and 15 months taken with an Opera Phenix confocal microscope
• These data were transformed with cellprofiler software using two pipelines:
  • the first pipeline performed projections and illumination corrections of 30 images in TIFF format containing a stack of piece of image to single image
  • the second pipeline enhanced the image, measured a number of features and extracted primary and secondary objects like cells and dendrites
• Data of this project:
  • 3243 files (summary-image data) contains 102 measures of image e.g. img_2-summary
  • 3243 files (secondary-dendrities) contains each containing on average 150 dendrities describing with 116 features img-2-secondary-dendrities

Tasks

• loading data, cleaning data, join summary image data with secondary dendrites
• preprocessing data with Standard Scaler
• feature engineering
  • calculate average area of dendrities in summary image data
• random under sample
• own cross validation

Classifiaction based on summary image data

Classification of 3256 pieces into healthy and sick classes based summary-image data

	LogisticRegression	MLPClassifier	Net
accuracy	0.747	0.704	0.748
recall	0.783	0.735	0.767
precision	0.757	0.700	0.775
f1_score	0.748	0.697	0.743