A multiresolution approach to automated classification of protein subcellular location images

BMC Bioinformatics

Table 1 Classification accuracy per class. Z, M and T stand for Zernike, morphological and texture features.

System	T	Weight.	Classification accuracy [%]
			M	T	Z	T, M	M, Z	T, Z	All
nMR	T ₁	NW	66.12	85.49	51.20	85.76	72.48	85.06	85.04
	T ₂	NW	66.12	85.76	51.20	86.64	72.48	85.78	86.24
	T ₃	NW	66.12	87.46	51.20	87.38	72.48	87.12	86.86
MRB	T ₃	OF	81.62	91.82	65.42	92.04	83.38	91.66	92.36
	T ₃	CF	81.48	92.32	65.84	92.62	83.58	92.34	92.54
MRF	T ₃	OF	84.92	94.72	65.82	94.64	86.80	94.74	94.52
	T ₃	CF	85.16	95.26	65.24	95.40	85.88	95.26	95.38

T₁ are the original Haralick texture features, T₂ are modified Haralick texture features from [16] and T₃ are our improved texturefeatures. nMR denotes the base system with no MR, MRB denotes MR basis classification and MRF denotes MR frame classification. OF denotes open-form weighting algorithm while CF denotes closed-form weighting algorithm. NW denotes no weighting as there is no MR block in front. Each entry is a number denoting the classification accuracy mean over a number of trials (different orderings of the images) for a given combination of feature sets. Note that the accuracy of nMR with features M is the same across the rows T1, T2, T3 since texture features are not involved in the classification when morphological features alone are used (similarly with Z, and M, Z). A subset of these results is shown pictorially in Figure 3. These results should be compared to the best previously obtained result (on the same data set) of 91.5% [15].

ISSN: 1471-2105