A Wavelet-based Method for Early Detection of Ovarian Cancer

Overview

The aim of this project is to describe a new method for detecting ovarian cancer. Based on the self-similarity of protein mass spectra, the new method method defines discriminatory descriptors that can be used to detect the presence of cancer. Wavelet spectra of protein mass spectra are used to assess self-similarity in the wavelet domain. A distance variance-based method is proposed as a novel method for computing wavelet spectra because it is more robust to outliers and noise, which typically present in protein mass spectra. The slope of the wavelet spectrum reflects the interplay between protein expression levels, which indicates the presence of cancer.

Dataset

The data used in the project available at the American National Cancer Institute Internet Repository. Two datasets namely Ovarian 4-3-02 (100 cases and 100 controls) and Ovarian 8-7-02 (162 cases and 91 controls) were used to validate the proposed modality. Each protein mass spectra consists of the intensities of 15,153 peptides defined by their mass-to-charge ratio (i.e. the ratio of molecular weight to electrical charge) (M/z).

Methods

Proposed Method

The proposed method assesses self-similarity of protein mass spectra in the wavelt-domain and define descriminatory descriptors to identify diagnostic features in protein mass spectra. The following concepts/methods are used in this method.

1. Wavelet Transform: a standard signal processing tool decomposes protein mass spectra into a hierarchy of resolutions convenient for assessing self-similarity.

2. Assessing Self-Similarity: self-similarity is a phenomenon that characterizes the stochastic similarity in protein mass spectra when viewed at different resolution scales. The following two methods are used to assess self-similarity:

   i). Standard Varaince Based Wavelet Spectra: consists of wavelet log-energies (logarithm of average of squared wavelet coefficients) as a function of resolution level. The regular decay (slope) in wavelet energies with a increase of resolution indicates self-similarity.

   ii). Distance Variance Based Wavelet Spectra: the standard variance wavelet spectra is sensitive to outliers and noise typically present in protein mass spectra. The distance variance-based method is proposed to assess self-similarity more precisely.

3. Rolling Window-based Approach: Every protein mass spectrum is divided into a set of data windows of size 1024, and the wavelet spectrum for each data window is calculated to determine the slope. As a result, a collection of slopes is generated, which are then used as discriminatory descriptors.

4. Fisher's criterion is used to select the most significant discriminatory descriptors.

Direct (Existing) Method

The Fisher's criteria is applied on the original protein mass spectra to select features with large differences in mean intensity values between cases and controls at each mass-to-chage ratio values.

Matlab Implementation

The following steps are explains procedure use to detect the presence of cancer on the basis of dataset Ovarian 4-3-02.

1. Load Data: DataRead_4_3_02.m reads data from the dataset Ovarian 4-3-02and then stores as ovarian12.mat.

2. Classifying Features: ComputeSlop_4_3_02.m computes wavelet spectra and estimates slopes under the standard variance and distance variance methods. The slopes are stored as SlopeDataset4_3_2.mat.

3. Classifiers: Three different classifires, Logistic Regression, Support Vector Machine, and K-Nearest Neighbor are used in OvarianJointclassify.m to perform classification.

4. Detection Of Cancer: Demo_Dataset_4_3_03.m performs the following three classifications:

   i). Self-Similarity based Method: using the descriptors computed from the standard variance and distance variance based methods.

   ii). Direct Method: using features selected by applying the Fisher's criteria on the protein mass spectra.

   iii). Joint Method: integrating features from both the self-similarity based method and the direct method.

Note: The same procedure is perfromed on the Ovarian 8-7-02 dataset using the codes named with 8-7-02.

Additional Evaluations

The MatchWindowFeature_4_3_2.m and MatchWindowFeature_8_7_2.m provide information about the location of the most significant descriptors in the protein mass spectra. The significant features correspond to the proposed method and direct method are displayed as regions (due to window-based method) and dashed lines on the protein mass spectra, respectively.

Outcomes

1. The proposed modality contributes to improvement in diagnostic performance for early detection.
2. Compared to the existing Ovarial cancer detection methods, the proposed modality has improved generalizability and reproducibility as it requires minimal pre-processing!

Note: Interested readers are encouraged to read the original manuscript.