HeLa Image Collection, Tom Macura
This is a summary of the analysis the Murphy group did using the HeLa cell experiment over the last few years.
R. F. Murphy (2004) Automated Interpretation of Protein Subcellular Location Patterns: Implications for Early Cancer Detection and Assessment. Annals of the New York Academy of Science 1020:124-131
The is a collection of images of HeLa cells that were labeled with anti-bodies against protein markers for various organellas: '(A)' ER protein, '(B)' Golgi protein giantin, '(C)' Golgi protein GPP130, '(D)' Lysosomal protein LAMP2, '(E)' mitochondrial protein, '(F)' nucleolar protein nuclelon, '(G)' actin labeled with rhodamine-phalloidin, '(H)' transferrin receptor, '(J)' cytoskeletal protein tubulin, '(K)' DNA labeled with DAPI. See Figure 1.
Figure 1: HeLa Dataset representative images
Automated Classification
For feature extraction they used Zernike moment features, Haralick texture features, morphological features, edge features, and hull features. They used stepwise discriminant analysis for optimal feature subset selection. For classification they used a back propagation neural network with a single layer of 20 hidden units over 10 cross-validation trials. Results are summarized in Figure 2. The average classification rate was 88%.
Figure 2: Results of Automated Classification on HeLa Dataset
Using more 2D features (e.g. Skelton Features, Gabor, and Daubechies 4 wavelets) and 3D features, they achieved an average accuracy of 95.8% with automated classification of the 3D HeLa image collection.
Human Visual Classification
"An automated system should recognize subtle differences in protein patterns that are not readily distinguishable by visual examination". To evaluate this hypothesis, they tested the ability of a human observer to learn this task. Results are summarized in Figure 3. The average classification rate was 83%.
Figure 3: Results of Manual Classification
"Of particular importance is the observation that two Golgi proteins included in the set could be distinguished with over 86% accuracy by automatic classification even though they could not be distinguished beyond random accuracy by visual examination." This highlights that subcellular proteins can be recognized with sensitivity and specificity higher than human observation with Murphy's automated classifier.
Plurality Voting
The results from Figure 2 are for classifying single cells. Based on the assumption all cells on a slide belong to the same class, the 'plurality voting' scheme makes the prediction for the whole set based on the class that had the most cells. Using this scheme, average accuracy increased from 83% to greater than 98%.
3D Features
They also collected a dataset of 3D HeLa images covering the same patterns as in the 2D data set. Using only 3D morphological features, these images could be classified with an average accuracy of 91%. This was 5% higher than obtained by 2D classification using only the central slice from each 3D image.
Comparision of Cell Populations
An important problem in fluorescence microscopy is determining whether the protein pattern changes in response to some treatment. More generally, this task can be described as determining whether two sets of images represent statistically different patterns. The Murphy group has developed a system ('SImEC for Statistical Imaging Experiment Comparator') that performs rigorous statistical comparision of image sets.
Using the Hotelling T2 test and other statistics, they concluded that the features they extract can be used to create a statically sound method for comparing subcellular protein distributions.
