By Ayoush Srivastava
The earliest record of cancer was found in a trauma surgery textbook from ancient Egypt; it described that for the disease, “there is no treatment” (ACS Medical Content, Editorial, and News Staff, 2014). Thousands of years later, cancer remains a formidable force, causing one-sixth of deaths across the globe (Zhu et al., 2020). While novel therapies are being developed to decrease the disease’s lethality, predicting the course of a patient’s cancer and their likelihood of surviving remains a vital task for clinicians (Zhu et al., 2020). Although classical statistical approaches – such as Cox proportional hazards (Cox-PH) – remain viable, deep learning’s sudden and conspicuous emergence proves to be a promising direction towards improving our ability to prognose a patient’s cancer.
Deep learning, a branch of machine learning, broke onto the AI scene a few years ago. Its popularity stems from one fact: the more data the model is trained with, the greater the model’s performance. Although the theory behind this technology was developed in the 1980s (Mathworks.com, 2019), the unprecedented growth of computational power and data captured by cellular and other devices has allowed for myriad practical applications (Zhu et al., 2020). Its function relies on mimicking neurons; these neurons can be arranged to form layers and can intake training data and output predictions. These layers can then be stacked upon each other to form different types of neural networks depending on the task at hand. For example, convolutional neural networks (CNN) are typically applied to image data, while recurrent neural networks (RNN) are applied to text and sequence data (Zhu et al., 2020).
Due to the rise of open-source databases for multi-omic data like the Cancer Atlas Genome (TCGA) and the Gene Expression Omnibus (GEO), clinical data has become increasingly accessible to test and train deep learning algorithms (Zhu et al., 2020). This has allowed researchers to create more comprehensive and accurate models for cancer prognosis. Various research groups have created neural network models and have applied them to datasets that vary in sample size, the patient’s type of cancer, and the multi-omic data available. Despite this variability among datasets, these experimental models can be classified into three categories: neural networks with no feature extraction; neural networks with feature extraction from gene expression data; and CNNs (Zhu et al., 2020).
For neural networks with no feature extraction, studies have shown that their performance is comparable to conventional statistical methods. For example, Cox-nnet, a neural network that takes gene expression data from TCGA to predict the patient’s survival time using Cox regression, achieved similar or better performance than conventional methods like Cox-PH (Zhu et al., 2020). Furthermore, another model named DeepSurv performed survival analysis using patient’s clinical data, specifically breast cancer data from METABRIC and GBSG as the input. By applying common hyperparameters for normalizing datasets, Katzman et al. were able to achieve more accurate results than a Cox-PH analysis (Zhu et al., 2020).
For neural networks with feature extraction from gene expression data, researchers often applied dimension reduction techniques to normalize and prepare the dataset for training the model; this is due to the high dimensionality and small sample size of health data (Zhu et al., 2020). To counteract this, Sun et al. applied a technique called minimum redundancy maximum relevance (mRMR) to reduce the dimensionality of gene expression and copy number alternation data for predicting breast cancer prognosis (Zhu et al., 2020). Their neural network model was able to outperform other machine learning methods like random forest, logistic regression and support vector machine. It should be noted, however, that the dataset used for training the model was unbalanced, meaning that the model’s performance may not carry over to new datasets. Other research groups applied feature extraction from multi-omic data using algorithms or by applying their expertise of biomedicine to track biological pathways (Zhu et al., 2020); however, imbalanced data reigned a common challenge along these attempts.
Convolutional neural networks (CNNs) have found great success across the medical field, being used to detect lymph node metastasis, classify skin lesions, and much more (Yamashita et al., 2018). In the field of cancer prognosis, CNNs have been tested on classifying cancerous tissue to predict survival of patients or extract image features for use in machine learning models (Zhu et al., 2020). One example is ResNet50, a 50-layer pre-trained CNN model that can identify a gene promoter in the brain associated with a higher survival rate from glioblastoma multiforme using MRI images (Zhu et al., 2020). This model was able to achieve an accuracy rate of about 95% (Zhu et al., 2020). Research groups have also trained CNNs to extract features from images to be used for other models. For example, Wang et al. trained a CNN using CT-based images of high-grade serous ovarian cancer to extract image features to help build a Cox-PH model, showcasing the flexibility CNNs offer for either classifying medical scans or extracting image features for further use in other models (Zhu et al., 2020).
However, many challenges remain to be resolved before neural network models can be implemented in the clinical field. One challenge is the lack of patient data for training. Since a neural network’s performance depends on the amount of data available for training, a small dataset can cause the model to overfit, preventing the model from performing optimally on an unassociated dataset (Zhu et al., 2020). Another challenge is the absence of a substantial database to store patient data. Since clinical data is sensitive and private to the patient, data security is a leading concern (Zhu et al., 2020). Furthermore, patient records can be imbalanced or noisy; for example, high mortality cancers will tend to have less survivors while some records may have missing fields of data, reducing the overall performance of the model (Zhu et al., 2020). Finally, researchers with expertise in deep learning and biomedical research are few and far between, but their expertise is required in labeling, organizing, and annotating medical data to render it usable for training deep learning models (Zhu et al., 2020).
In conclusion, there is great promise behind the application of deep learning towards cancer prognosis, supplying clinicians with a new avenue for predicting the course of a patient’s condition. However, multiple obstacles – like the lack of usable training data and absence of experts in biomedical research and deep learning – will need to be overcome in order to implement these models in the clinical setting.
ACS Medical Content, Editorial, and News Staff (2014). Early History of Cancer. [online] Cancer.org. Available at: https://www.cancer.org/cancer/cancer-basics/history-of-cancer/what-is-cancer.html.
Mathworks.com. (2019). What Is Deep Learning? | How It Works, Techniques & Applications. [online] Available at: https://uk.mathworks.com/discovery/deep-learning.html.
Zhu, W., Xie, L., Han, J. and Guo, X. (2020). The Application of Deep Learning in Cancer Prognosis Prediction. Cancers, 12(3), p.603.
Ching, T., Zhu, X. and Garmire, L.X. (2018). Cox-nnet: An artificial neural network method for prognosis prediction of high-throughput omics data. PLOS Computational Biology, 14(4), p.e1006076.
Katzman, J.L., Shaham, U., Cloninger, A., Bates, J., Jiang, T. and Kluger, Y. (2018). DeepSurv: personalized treatment recommender system using a Cox proportional hazards deep neural network. BMC Medical Research Methodology, 18(1).
Sun, D., Wang, M. and Li, A. (2019). A Multimodal Deep Neural Network for Human Breast Cancer Prognosis Prediction by Integrating Multi-Dimensional Data. IEEE/ACM Transactions on Computational Biology and Bioinformatics, [online] 16(3), pp.841–850. Available at: https://ieeexplore.ieee.org/abstract/document/8292801 [Accessed 25 Feb. 2021].
Huang, Z., Zhan, X., Xiang, S., Johnson, T.S., Helm, B., Yu, C.Y., Zhang, J., Salama, P., Rizkalla, M., Han, Z. and Huang, K. (2019). SALMON: Survival Analysis Learning With Multi-Omics Neural Networks on Breast Cancer. Frontiers in Genetics, [online] 10, p.166. Available at: https://pubmed.ncbi.nlm.nih.gov/30906311/ [Accessed 25 Feb. 2021].
Hao, J., Kim, Y., Kim, T.-K. and Kang, M. (2018). PASNet: pathway-associated sparse deep neural network for prognosis prediction from high-throughput data. BMC Bioinformatics, 19(1).
Yamashita, R., Nishio, M., Do, R.K.G. and Togashi, K. (2018). Convolutional neural networks: an overview and application in radiology. Insights into Imaging, 9(4), pp.611–629.
Korfiatis, P., Kline, T.L., Lachance, D.H., Parney, I.F., Buckner, J.C. and Erickson, B.J. (2017). Residual Deep Convolutional Neural Network Predicts MGMT Methylation Status. Journal of Digital Imaging, [online] 30(5), pp.622–628. Available at: https://pubmed.ncbi.nlm.nih.gov/28785873/ [Accessed 25 Feb. 2021].
Wang, S., Liu, Z., Rong, Y., Zhou, B., Bai, Y., Wei, W., Wei, W., Wang, M., Guo, Y. and Tian, J. (2019b). Deep learning provides a new computed tomography-based prognostic biomarker for recurrence prediction in high-grade serous ovarian cancer. Radiotherapy and Oncology: Journal of the European Society for Therapeutic Radiology and Oncology, [online] 132, pp.171–177. Available at: https://pubmed.ncbi.nlm.nih.gov/30392780/.