Automated Classification of Osteosarcoma and Benign Tumors using RNA-seq and Plain X-ray

Abstract

Osteosarcoma is a prominent bone cancer that typically affects adolescents or people in late adulthood. Early recognition of this disease relies on imaging technologies such as x-ray radiography to detect tumor size and location. This paper aims to differentiate osteosarcoma from benign tumors by analyzing both imaging and RNA-seq data through a combination of image processing and machine learning. In experimental results, the proposed method achieved an Area Under the Receiver Operator Characteristic Curve (AUC) of 0.7272 in three-fold cross-validation, and an AUC of 0.9015 using leave-one-out cross-validation.

I. INTRODUCTION

More than 1,900 cancer-related deaths occur in children and adolescents every year in the United States. From 2007–2010, bone cancers accounted for 5% of cancer mortality in children less than 15 years, which increased to 15% in those aged 15–19 years [1]. Osteosarcoma is the most common primary bone malignancy; it has a bimodal distribution in age, with the first peak occurring in adolescence, and the second in older adulthood. [2]. A primary method of tumor diagnosis is x-ray radiography, which is used to locate the tumor and assess its qualities. Some typical features of osteosarcoma tumors in x-ray include a sun-burst appearance, periosteal lifting with the formation of Codman’s triangle, and the creation of new bone in soft tissue, though there are variations depending on the specific type of osteosarcoma [3]. A timely diagnosis is needed in order to increase the chance of patient survival and reduce the risk of metastasis.

Previous work exists related to incorporating multimodal data for osteosarcoma classification. In Shen et al. [4], x-ray imaging features coupled with metabolomic data were used to create a random forest (RF) classification model that achieved an Area Under the Receiver Operator Characteristic Curve (AUC) of 0.94 (standard deviation 0.054). An advantage of RF is that it does not require the number of features to be smaller than the number of samples used for training. Studies such as [5], [6], have used RFs successfully in such instances. Wu et al. [5] compared the performance of RF with other classifiers in discriminating normal from ovarian cancer serum samples with mass spectrometry data, finding that RF produced the least variation in prediction error when compared to other tree-based classifiers. Lee et al. [6] similarly found RF to be the strongest tree-based classifier when using microarray data as model input.

In this paper, we propose an automated method for classification of benign and osteosarcoma tumors using RNA-seq and x-ray images. First, feature extraction methods to obtain salient information from raw RNA-seq data and x-ray images are described. Given the high number of features obtained, chosen methods to reduce the number of features are introduced. Lastly, the performance of the proposed method on an x-ray image and RNA-seq dataset collected by collaborators at Shanghai Jiao Tong University (SJTU) is discussed. The experimental results suggest that automated classification of benign and malignant tumors can be used to enhance computer-aided diagnosis.

II. MATERIALS AND METHODS

A. Dataset

Demographics are provided in Table I. All of the samples contain both RNA-seq data and plain x-ray images by collaborators at SJTU. The RNA sequencing libraries were sequenced in 200bp paired-end mode using the Illumina HiSeq system (Illumina, San Diego, CA, USA). Sample collection for this study was conducted following a Shanghai General Hospital IRB-approved protocol after all participants signed written informed consent. A schematic diagram depicting the overall classification system in presented in Fig. 1.

Table I:

Characteristics of Patients

Tumor Type	Mean Age (SD)	Sex
Benign	40.33 (12.48)	5 Male, 1 Female
Osteosarcoma	22.73 (16.14)	4 Male, 7 Female

Fig. 1:

A schematic diagram of the proposed classification system.

B. Image Processing

1). Segmentation:

Imaging data consisted of plain x-ray DICOM images of the tumor and surrounding bone. Images were processed following the methods defined in [4]. A semi-automated segmentation method, Graph Cut with Lazy Snapping [7], was used to segment images into tumor and affected bone regions before feature extraction. With Graph Cut, the user specifies foreground and background seeds to serve as hard constraints for the segmentation. Graph Cut then assigns labels to each node x_i (such as 1 for foreground and 0 for background) to minimize the Gibbs Energy E(X) on a graph with respect to the user-defined hard constraints.

Before assigning foreground and background labels, the watershed algorithm pre-segments the graph into superpixels, resulting in a graph $\mathcal{G}=\langle \mathcal{V},\mathcal{E}\rangle$, where $\mathcal{V}$ represents all the small regions, and $\mathcal{E}$ represents all the arcs that connect them. The energy function E(X) to minimize is defined as

$$E(X)=\sum _{i\in \mathcal{V}}{E}_1\left(x_i\right)+\lambda \sum _{i,j\in \mathcal{E}}{E}_2\left(x_i,x_j\right),$$ (1)

where E₁(x), the likelihood energy, measures the cost of labeling node i as x_i, and E₂(x_i,x_j), the prior energy, the cost of labeling connected nodes i and j as x_i and x_j.

Fig. 2 shows the Graph Cut with Lazy Snapping method applied to an x-ray image. Fig. 2a is the original x-ray image of the tumor and bone region. Fig. 2b shows the user-defined foreground seeds (green dots) and background seeds (red dots) to select the tumor regions. Fig. 2c displays the foreground labels that minimize the Gibbs Energy. Fig. 2d is the final mask for the tumor region.

Fig. 2:

Graph Cut with Lazy Snapping applied to tumor segmentation. [4]

2). Feature Extraction:

The process for extracting features is detailed in previous work [4] and is illustrated in Fig. 3. There are features collected from both the bone and tumor segmentations (tumor border clarity, distance to joint), as well as from the tumor segmentation alone (texture features, morphological features). Table II lists all features extracted from the x-ray images.

Fig. 3:

The image processing techniques utilized for feature extraction.

Table II:

Features from Image Processing

Morphological Features	Mean, Standard Deviation, Entropy, Kurtosis, Skewness, Convex Area, Eccentricity, Perimeter, Major Axis Length, Minor Axis Length, Solidity
Texture Features	Contrast, Correlation, Energy, Homogeneity from gray-level co-occurrence matrix; Mean, Standard Deviation, Kurtosis, Skewness, and Entropy from Approximation, Horizontal Detail, Vertical Detail, and Diagonal Detail Coefficients
Other Features	Distance to Joint, Tumor Border Clarity

C. RNA-seq Processing

A summary of the steps taken to process RNA-seq data for analysis is presented in Fig. 4. First, the quality of all of the samples to be included in the analysis was verified (Section II-C.1). Then, data were aligned to the human reference genome, and raw counts for the aligned reads were obtained (Section II-C.2). After this process, the count data were normalized (Section II-C.3) and reduced (Section II-C.4) before being input to the RF model.

Fig. 4:

Processing RNA-seq data to obtain counts.

1). Data Quality Assessment:

Before aligning RNA-Seq data against the reference genome, FastQC version 0.11.8 [8] was used to assess the quality of the raw RNA-seq data. The raw FASTQ files were used as input to FastQC, which was run using default parameters. For each sample, FastQC generated an HTML report, showing scores and graphics that illustrated quality metrics such as per-base sequence quality, GC content, duplicate, and over-represented sequence. We examined these reports before proceeding to the alignment step. The overall sequence quality was deemed acceptable, and thus, no samples were removed from the analysis, nor was adapter trimming applied.

2). Genome Alignment and Generation of Count Data:

Using STAR [9] with default parameters, a genome index was generated from the current human reference genome (hg38) and its corresponding General Transfer Format (GTF) annotation file. The reference genome and GTF files were obtained from the UCSC website [10], using Gencode V29 [11]. STAR was then used to map the raw RNA-Seq reads against the indexed reference genome. The maximum number of multiple alignments allowed for a read was set to 20. The resultant BAM file was sorted by coordinate as required by QoRTs [12]. All samples achieved a ratio of uniquely mapped reads greater than 76%. The aligned RNA-Seq data were processed via the Java utility in QoRTs [12]. The processed data were then read into R using the QoRTs library to generate count files. These counts, after performing the normalization and feature reduction described in the following sections, serve as RNA-seq features for the classification model.

3). Normalization of RNA-seq Count Data:

With RNA-seq data, counts can vary from sample to sample, and should therefore be normalized. The normalization method used in this paper follows the example of [13], which relies on DESeq2’s [14] normalization method. DESeq2 takes raw counts as input, then uses a median ratio equation [14], [15] to normalize the counts across samples. For each sample, DESeq2 calculates the size factor, then divides the sample’s raw counts by its size factor. For n transcripts and m samples, the size factor ${\widehat{s}}_j$ for sample j is calculated as

$${\widehat{s}}_j=\mathit{\text{media}}{n}_i\frac{k_{ij}}{{\left({\prod }_{v=1}^m{k}_{iv}\right)}^{1/m}},$$ (2)

with k_ij the raw count for transcript i in sample j. Any transcript having a count of zero was excluded from this calculation, following the method from DESeq2 [15], [16].

The training data were normalized using the size factors from (2), which were also used to derive the size factors $\widehat{s_{j^{\ast }}}$ for each sample j* in the test set, shown in (3).

$$\widehat{s_{j^{\ast }}}=\mathit{\text{media}}{n}_i\frac{k_{i{j}^{\ast }}}{{\left({\prod }_{v=1}^m{k}_{iv}\right)}^{1/m}}$$ (3)

In this manner, the test set is also normalized by the data in the training set. This approach is justified, for if the original equation from [15] was employed, i.e., using all of the data available for normalization, the training samples would be normalized by both themselves as well as the test samples, which would lead to data leakage. Likewise, the test samples would also be normalized by themselves, leading to overfitting and data leakage.

4). RNA-seq Feature Reduction:

Due to the large number of features (counts of transcripts) in our RNA-seq dataset, we produced a list of genes of interest. We limited our inclusion of RNA-seq transcript counts to those which could be associated with the genes of interest. A previous literature search identified 211 genes either whose upregulation is associated with tumorigenesis and metastasis, or whose downregulation is associated with tumorigenesis and metastasis. This feature reduction step was performed to reduce potential noise being added to the classification model, and to limit the RNA-seq features to those which could be biologically relevant.

D. Machine Learning

After performing image processing and selecting the RNA-seq features of interest, the two data types were combined into one dataset. In this manner, every observation consisted of one individual, containing both features from image processing and counts from transcripts of interest from RNA-seq.

To create the classification of osteosarcoma or benign tumor, RF [17] was used. In each model, the number of trees used was 50, the square root of the total number of features was selected for each decision split, and the minimum leaf size was 1. RF models were constructed using both three-fold cross-validation and leave-one-out cross-validation.

1). Three-Fold Cross-Validation:

For three-fold cross-validation (CV), three folds of observations were created, with each fold containing 2 benign observations and 3–4 of osteosarcoma. The observations in each fold were selected randomly, and there was no overlap among each of the folds. Two folds were used to train the RF model, and the third fold was used to test performance. This was repeated, utilizing each fold once as the test fold. Before data were input to the model, principal component analysis (PCA) was applied to the training folds to reduce the feature space. The test fold was scaled by the training folds, and the coefficients from PCA applied to the training folds were used to yield the principal components (PCs) for the test fold, in order to prevent data leakage.

2). Leave-One-Out Cross-Validation:

For this method, all observations except one were used as a training set for the RF model, and the last observation was used to test performance. This was repeated so that each observation was used once for testing. Before data were input to the model, PCA was applied to the training set to reduce the feature space. The test observation was scaled by the training data, and the coefficients from PCA applied to the training set were used to yield the PCs for the test set.

III. RESULTS & DISCUSSION

In this study, a RF model was created using features extracted from RNA-seq and x-ray image data to classify a given tumor as benign or osteosarcoma. A dataset was developed, containing imaging features as well as RNA-seq counts from transcripts associated with genes from a relevant literature search. Before training each model, PCA was performed to reduce the feature space of the imaging features combined with transcript counts in the training set. Three-fold and leave-one-out CV were used to assess performance. Performing three-fold CV with RF trained on only one PC achieved an AUC of 0.7272. The performance results of three-fold CV with one PC are displayed in Table III. Performing leave-one-out CV achieved AUC of 0.9015 when 2 PCs were used to train the model. The performance results are displayed in Table IV.

Table III:

Average Performance of the Random Forest Model with Three-Fold Cross-Validation

PCs	AUC	F1	Sensitivity	Specificity	Accuracy
1	0.7272	0.8462	1	0.3333	0.9647

Table IV:

Average Performance of the Random Forest Model with Leave-One-Out Cross-Validation

PCs	AUC	F1	Sensitivity	Specificity	Accuracy
1	0.4470	0.8148	1	0.1667	0.7059
2	0.9015	0.9091	0.9091	0.8333	0.8824

One major concern when performing this study was limiting the number of variables to train the model. Thousands of transcripts were aligned to the genome. Thus, PCA was used to capture the variance explained by those features, while limiting the number of features exposed to the model.

Increasing the number of PCs increases the amount of information being provided to the model, so it is expected that additional PCs lead to higher AUC, as shown in Table IV. However, performance in both three-fold and leave-one-out CV did not significantly improve after adding more PCs than those displayed, which may indicate that a large amount of information introduced is noise.

Limitations of this study include small sample size and class imbalance. The small number of observations in this study reduces the ability to assess the generalizability of this classification method. For example, because there were only 6 benign cases, a difference of 1 incorrect classification could result in large changes in specificity. Additionally, there are almost twice as many osteosarcoma cases as benign tumor cases in our dataset. Despite these limitations, the results suggest that the models created in this study may be used to assist clinicians in accurately distinguishing between benign and malignant tumors in target patient populations.