A multiresolution approach to automated classification of protein subcellular location images

Abstract

Background: Fluorescence microscopy is widely used to determine the subcellular location of proteins. Efforts to determine location on a proteome-wide basis create a need for automated methods to analyze the resulting images. Over the past ten years, the feasibility of using machine learning methods to recognize all major subcellular location patterns has been convincingly demonstrated, using diverse feature sets and classifiers. On a well-studied data set of 2D HeLa single-cell images, the best performance to date, 91.5%, was obtained by including a set of multiresolution features. This demonstrates the value of multiresolution approaches to this important problem.

Results: We report here a novel approach for the classification of subcellular location patterns by classifying in multiresolution subspaces. Our system is able to work with any feature set and any classifier. It consists of multiresolution (MR) decomposition, followed by feature computation and classification in each MR subspace, yielding local decisions that are then combined into a global decision. With 26 texture features alone and a neural network classifier, we obtained an increase in accuracy on the 2D HeLa data set to 95.3%.

Conclusion: We demonstrate that the space-frequency localized information in the multiresolution subspaces adds significantly to the discriminative power of the system. Moreover, we show that a vastly reduced set of features is sufficient, consisting of our novel modified Haralick texture features. Our proposed system is general, allowing for any combinations of sets of features and any combination of classifiers.

Figure 1

Basic multiresolution block. Top: Two-channel analysis filter bank. The filter h is a highpass filter and g is a lowpass filter. Bottom: A 2-level filter bank decomposition of actin. If the original image is of size N × N, the ones in the middle are of sizes N/2 × N/2 and the ones on the right are of sizes N/4 × N/4. Each branch has either the lowpass filter g or the highpass filter h followed by downsampling by 2 as in the top figure. Filtering and sampling are performed along the horizontal direction (rows) followed by the same operations along the vertical direction (columns).

Figure 2

Multiresolution (MR) classification system. The generic classification system (GCS) consists of feature extraction followed by classification (inside the dashed box). We add an MR block in front of GCS and compute features in MR subspaces (subbands). Classification is then performed on each of the subbands yielding local decisions which are then weighed and combined to give a final decision.

Figure 3

Pictorial representation of classification accuracy results. The diagram shows results from Table 1 for those sets involving T₃, namely (T₃), (T₃, M) and (T₃, M, Z). Diamond markers represent the nMR system (no MR block), circles represent the MRB system (MR bases, no redundancy) and squares represent the MRF system (MR frames, redundancy). Filled markers denote the closed-form weighting algorithm (CF), while empty ones denote the open-form weighting algorithm (OF). The following trends are noteworthy: (a) Introducing MR (both MRB and MRF) significantly outperforms nMR, thus demonstrating that classifying in MR subspaces indeed improves classification accuracy. (b) MRF outperform MRB. (b) For the two versions of the weighting algorithm, open form and closed form, the closed-form algorithm slightly outperforms the open-form one. (d) The trend in each case is almost flat across various feature set combinations, indicating that the texture set T₃ alone (26 features) is sufficient for high classification accuracy.