Cell Detection from Microscope Images Using Histogram of Oriented Gradients
Tuomas Tikkanen · Tampere University Institutional Repository (Tampere University) · 2014
In this master's thesis, cell detection from bright-field microscope images is studied using Histogram of Oriented Gradients (HOG) features with Support Vector Machine supervised learning classifier. The proposed method is trained iteratively by finding hard examples. The performance of the method is evaluated using 16 training and 12 test images with altogether 10736 PC3 human prostate cancer cells. The cell detection accuracy is assessed with Receiver Operating Characteristic curves, F1-score and Bivariate Similarity Index. Decision between true positive and false positive detection is made using PAS-metric. The experiments consider various parameters and their effect on performance. It is shown that using hard examples in the training phase increases the level of generalization of the model considerably. ROC AUC reaches an excellent value of 0.98 after iterative training. When SVM threshold is varied for each image in the testing phase, F1-score averaged over the peak F1-scores of each image reaches a high value of 0.85. The most suitable combinations of HOG descriptor parameter values are presented. All in all, results indicate that HOG can be successfully applied to bright-field microscope images of PC3 prostate cancer cells taken on subsequent days, which results in a growth curve that favorably agrees with manual counts. The implemented cell detection framework outperforms humans in terms of consistency, objectivity and speed.