Deep learning-extracted CT imaging phenotypes predict response to total resection in colorectal cancer

Abstract

Background

Deep learning surpasses many traditional methods for many vision tasks, allowing the transformation of hierarchical features into more abstract, high-level features.

Purpose

To evaluate the prognostic value of preoperative computed tomography (CT) image texture features and deep learning self-learning high-throughput features (SHF) on postoperative overall survival in the treatment of patients with colorectal cancer (CRC).

Material and Methods

The dataset consisted of 810 enrolled patients with CRC confirmed from 10 November 2011 to 10 February 2018. In contrast, SHF extracted by deep learning with multi-task training mechanism and texture features were extracted from the CT with tumor volume region of interest, respectively, and combined with the Cox proportional hazard (CoxPH) model for initial validation to obtain a RAD score to classify patients into high- and low-risk groups. The SHF stability was further validated in combination with Neural Multi-Task Logistic Regression (N-MTLR) model. The overall recognition ability and accuracy of CoxPH and N-MTLR model were evaluated by C-index and Integrated Brier Score (IBS).

Results

SHF had a more significant degree of differentiation than texture features. The result is (SHF vs. texture features: C-index: 0.884 vs. 0.611; IBS: 0.025 vs. 0.073) in the CoxPH model, and (SHF vs. texture features: C-index: 0.861 vs. 0.630; IBS: 0.024 vs. 0.065) in N-MTLR.

Conclusion

SHF is superior to texture features and has potential application for the preoperative prediction of the individualized treatment of CRC.

Keywords

Self-learning high-throughput features deep learning overall survival colorectal cancer computed tomography

Introduction

Colorectal cancer (CRC) is one of the most common malignant tumors (1). Due to its high incidence and mortality rates, CRC constitutes a significant portion of healthcare expenditure and a heavy burden on families and communities. Therefore, more accurate prediction of the survival rate of patients with cancer, better clinical diagnosis, and treatment decisions, including the choice of treatment methods, treatment timing, and follow-up visits, have a great impact on the treatment cost and treatment effect (2).

Different treatment options, tumor heterogeneity, and genetics can affect overall survival (OS). The progress of biological and genomic technology has improved the estimation of survival rate (3). Precision medicine takes into account individual differences in prevention and corresponding diagnostic and treatment strategies (4). Depending on the location and extent of the disease, the appropriate surgical treatment is chosen, including local surgery and more invasive procedures (5). Tumor heterogeneity, a recognized feature of malignancy, is also an important prognostic factor (6). However, the inability to obtain comprehensive information about heterogeneous tumors remains a limitation of these invasive methods (7). A large amount of objective and dense information is collected in medical images and computed tomography (CT) images can help identify important tumor characteristics (8). Extracting high-throughput quantitative descriptors from CT research and conducting a non-invasive analysis (9) provides insights into the individualized treatment practice of tumor detection, subtype classification, and treatment response evaluation (10). The description of image features will usually require physical markers with the ultimate area of interest to be labeled in the CT image. This work will require an expert, usually a radiologist, to label the features.

"Radiomics" can be broadly defined as the use of high-throughput computer technology to analyze high-dimensional data of medical images (11). Traditional methods of image analysis with human-defined features have been used in the radiology and informatics communities and have shown significant promise in finding useful biomarkers of health, particularly in cancer subclassification (12). Deep learning has developed from the computer science community and rapidly overtaken more traditional methods in many vision tasks, such as image recognition and segmentation (13).

The aim of the present study was to investigate the value of preoperative CT image texture features and deep learning self-learning high-throughput features (SHF) on postoperative OS in the treatment of patients with CRC.

Material and Methods

Approval was obtained from the institutional review board, and the requirement of written informed consent was waived for this retrospective study.

Patients

This retrospective study screened the sample by retrieving medical records from our hospital from 10 November 2011 to 10 February 2018. Preoperative CT scans were performed in each case, and clinical follow-up was performed 5–108 months after the scan. Surgery was performed in all cases after the CT scan and 810 patients were finally included. Fig. 1 shows the specific inclusion and exclusion criteria.

Fig. 1.

Flow chart of study cohort inclusion and exclusion criteria.

Image acquisition

Before CT examination, an 8-h fasting requires for preparation on the patients. The patients were in a supine position and all scanning was completed while the patients held their breath. Scans were performed on a SOMATOM Sensation 64 (Siemens Medical Solutions, Forchheim, Germany) or GE Optima CT660 scanner (GE Medical Systems, Milwaukee, WI, USA) with the following parameters: slice thickness and spacing = 5 mm; tube voltage = 120 kV; tube current = 100 mA; matrix = 512 × 512; and pitch = 0.984. After the unenhanced image was scanned, 90 mL of contract agent ioversol (320 mgI/mL) was injected via an automatic high-pressure syringe through the median cubital vein with an injection speed of 2.5 mL/s. Following a 70-second delay, enhanced images of the portal venous phase were obtained.

Tumor area labeling and feature extraction

The primary lesion and the largest peripheral lymph node volumes of interest (VOI), including associated blood vessels and lymph nodes, adipose tissue and rectal mesentery, were labeled by two radiologists with relevant experience using ITK-SNAP version 3.6.0 (http://www.itksnap.org) software and subsequently validated by senior radiologists with over 12 years of experience (14). Selected slices from each patient were contour-labeled for the region of interest (ROI) of the primary tumor mass. Ultimately, 4319 sections from 76,873 DCM samples from 810 patients were labeled.

The traditional texture feature extraction used special medical imaging software package Pyradiomics (https://pyradiomics.readthedocs.io). The extracted texture features include first order, 2D or 3D shape, gray level cooccurence matrix, gray level size zone matrix, gray level run length matrix, and gray level dependence matrix.

For SHF extracted based on the deep learning method with multi-task training mechanism, we first directly extracted the feature matrix in the labeled area image, and the features of each unit were stored in the feature matrix (32 × 32 × 32). Finally, the data were integrated to obtain the data size of N × 32 × 32 × 32, where N represents the total number of individuals. Its features were put into the deep high throughput feature self-learning network with multi-task training mechanism for feature selection; its modular network model structure is shown in Fig. 2.

Fig. 2.

Network structure diagram of deep high throughput feature self-learning network with multi-task training mechanism.

The model structure is based on an encoder and decoder. The encoder consists of three similar blocks. Each block includes convolution, linear rectifying function (RELU), and maximum pooling. Correspondingly, the decoder consists of three similar blocks. Compared with the encoder, deconvolution is used to replace maximum pooling. Finally, we get the feature information at the end of the connection layer as SHF. After obtaining the SHF, SoftMax is used as the loss function.

To achieve automatic feature extraction, the ROI is regarded as both the input and the ground truth. The corresponding loss function is defined as follows:

L_{reconstruction} = \frac{1}{n} \sum_{i = 1}^{n} (y_{i} - x_{i})^{2}

(1)

Where n represents the number of ROIs,

x_{i}

is the input, and

y_{i}

is the output.

To enhance the quality and clarity of the output in the reconstruction process, we introduced a perceptual loss to guide its learning. We obtained the feature maps in the middle layer of the encoder and decoder. The feature maps obtained from the same layer of encoder and decoder are defined as $ϕ_{i}^{e}$ and $ϕ_{i}^{d}$ , respectively. The specific calculation is as follows:

L_{perceptual} = \frac{1}{C_{i} \times H_{i} \times W_{i}} ϕ_{i}^{e} - ϕ_{i 2}^{d 2}

(2)

Where i represents the i -th layer,

C_{i}

H_{i}

, and

W_{i}

represent the channel, height, and width of feature map

ϕ_{i}^{e}

, respectively.

Feature dimensionality reduction and predictive model selection

The SHF features we obtained were high latitude and had some redundancy; thus, we had to perform a feature dimensionality reduction on them. To address this issue, the minimum absolute contraction and selection operator (Lasso) regression model, applicable to the regression of high-dimensional data, was used to select the most useful predictive features in the training dataset (15).

To assess the importance of the characteristics obtained by their different methods as prognostic variables on outcomes such as death or cancer recurrence, different survival validation models, Cox proportional hazard (CoxPH) and Neural Multi-Task Logistic Regression (N-MTLR), were chosen. The CoxPH, a standard survival model, is a semi-parametric model used to calculate the effect of its covariates on the risk of event occurrence. In addition, to further validate the stability and validity of our extracted features, we chose the more flexible N-MTLR, which uses a deep learning framework through a multilayer perceptron and does not rely on any CoxPH model assumptions.

The N-MTLR (16) is proposed to address the drawbacks of the CoxPH, such as the assumption that the ratio of the hazards for any two individuals is constant over time, all elements are assumed to be linearly related. The N-MTLR can be seen as a series of logistic regression models built on different time intervals so as to estimate the probability that the event of interest happened within each interval. In addition, it allows the use of neural networks to solve non-linear elements. We first built a survival time axis J and built a logistic regression model s on each interval. The density and survival functions can be defined as below.

Density function:

f = \frac{\exp (φ (\vec{x}) \cdot Δ \cdot \vec{Y})}{Z (φ (\vec{x}))}

(3)

Survival function:

S = \sum_{k = s}^{J} \frac{\exp (φ (\vec{x}) \cdot Δ \cdot \vec{Y})}{Z (φ (\vec{x}))}

(4)

With:

Z (φ (\vec{x})) = \sum_{j = 1}^{J} \exp (\sum_{j = 1}^{J} φ (\vec{x}))

(5)

Where

\vec{Y}

is the state of patient in each time interval,

Δ

is a triangular matrix, and

φ (\vec{x})

is the non-linear output of neural networks.

Decision curve analysis

Decision curve analysis (DCA) is a method of assessing the value of the information provided by a diagnostic test by considering the likely range of patient risk and benefit preferences without actually measuring these preferences for a particular patient (17). We use the R version 4.0.5 (https://www.r-project.org/) language to build their models and plot DCA curves to construct clinical models based on clinical data, traditional texture feature models extracted based on traditional methods, and SHF feature models extracted based on depth-based methods, respectively.

Statistical analysis

The inter-group test in Table 1, for continuous variables, was tested using a Mann–Whitney U test and discrete variables were tested using a Pearson's chi-square test. X-tile (https://medicine.yal-e.edu/lab/rimm/research/software/) software was used to obtain the RAD score cutoff value, which was used to determine the classification of high and low risk (18). We classified patients into high- and low-risk groups according to their RAD score cutoff values and plotted their Kaplan–Meier (KM) curves using IBM SPSS version 26 (IBM Corp., Armonk, NY, USA) (19). Cox regression was used for both the univariate and multivariate analyses. P values <0.05 were considered to be significant correlations. Texture features and SHF were extracted on the training set, respectively, and the survival of both was compared using CoxPH; KM curves were plotted separately. Further validation was obtained on the validation set.

Table 1.

Dataset clinical information.

	Total set (n = 810)	Training dataset (n = 600)	Validation dataset (n = 21)	P value
Age (range)	60 (19–90)	60 (29–90)	60 (19–86)	0.104
Age (years)				0.20
Elderly group (age ≥60)	494 (60.9)	359 (59.8)	135 (64.3)
Non-elderly group (age <60)	316 (39.1)	241 (41.2)	75 (35.7)
Sex				0.15
Male	479 (59.1)	346 (57.7)	133 (63.3)
Female	331 (40.9)	254 (42.3)	77 (36.7)
Tumor location				0.23
Rectum	411 (50.8)	312 (52)	99 (47.1)
Colon	399 (49.2)	288 (48)	111 (52.9)
Pathological grade				0.36
High or middle	578 (71.4)	423 (70.5)	155 (73.8)
Low	232 (28.6)	177 (29.5)	55 (26.2)
Histotype				0.006
Adenocarcinoma	689 (85.1)	522 (87.0)	150 (71.5)
Mucus histology	97 (12.0)	61 (10.2)	36 (17.2)
Others	24 (2.9)	17 (2.8)	7 (3.3)
Lymph node metastasis				0.90
Yes	294 (36.3)	217 (36.2)	77 (36.7)
No	516 (63.7)	383 (63.8)	133 (63.3)
Tumor size (cm)	4.633 ± 1.834	4.712 ± 1.683	4.554 ± 1.981	0.179
Stage				0.087
I	155 (19.1)	115 (19.2)	40 (19.1)
II	303 (37.4)	212 (35.3)	91 (43.3)
III	352 (43.5)	273 (45.5)	79 (37.6)

Values are given as n (%), mean ± SD or median (range).

Results

Population characteristics

The demographic and clinical characteristics of the 810 patients (479 men, 331 women; median age = 60 years; age range = 19–90 years) are shown in Table 1. There were 411 tumor sites in the rectum and 399 in the colon. The age of 60 years was the cutoff, dividing the patients into the elderly and non-elderly groups (60.9% vs. 39.1%; P = 0.004), and based on pathological grade, they were classified as high or middle and low differentiation (71.4% vs. 28.6%; P < 0.001). Histotype was divided into adenocarcinoma, mucus histology, and others (85.1% vs. 12.0% vs. 2.9%; P = 0.011). We partitioned the dataset in a ratio close to 3:1 (training:validation). The data for the 810 individuals were divided into 600 training datasets and 210 validation datasets, including specific clinical disaggregated statistics such as gender, pathological grade, and stage.

The endpoint of this study was OS, which was defined as the time from the date of the CT examination to death from any cause. When recording patients who were lost to follow-up before death, it is usually the time to the last recorded contact with the patient. The median OS of the entire sample was 57 months (95% confidence interval = 56–59). After Cox regression analysis, the results are shown in Table 2. In this case, when the multivariate regression analysis was performed, other blocks of relevant variables were added to their covariates, and the results showed that age, stage, and pathological grade were significantly associated with survival time in the multivariate analysis.

Table 2.

Results based on univariate and multivariate regression analysis.

	Univariate			Multivariate
Characteristic	HR	95% CI	P	HR	95% CI	P
Age
<60 years*	0.606	0.444–0.827	0.002	0.567	0.415–0.774	<0.001
Sex
Female	0.882	0.658–1.181	0.399
Stage			<0.001			<0.001
II	0.301	0.186–0.488		0.351	0.192–0.642
III	0.405	0.290–0.555		0.479	0.293–0.784
Location
Colon	0.909	0.761–1.087	0.295
Pathological grade
Low	2.953	1.600–5.450	0.001	2.907	1.576–5.364	0.001
Histotype			0.201
Mucus histology	5.791	0.811–41.361
Others	6.217	0.842–45.889
Lymph node metastasis
No	0.423	0.318–0.563	<0.001	0.812	0.518–1.273	0.364
Tumor size (cm)	0.896	0.719–1.117	0.329

*Median age is 60 years.

CI, confidence interval; HR, hazard ratio.

The effectiveness analysis of extracted features

The most useful predictive features were selected to be placed into the CoxPH model to obtain radiomics scores (RAD score). The RAD score was assigned to each patient by combining the selected features linearly weighted by their respective coefficients. The deep learning method represents SHF, and the traditional method represents texture features. The results of the validation data on the CoxPH model are shown in Table 3. N-MTLR and CoxPH used the Concordance Index (C-index) and Integrated Brier Score (IBS) as performance indicators of the evaluation model (16). In the CoxPH model, the results are as follows: SHF versus texture features: C-index: 0.884 versus 0.611; IBS: 0.025 versus 0.073. The C-index evaluates the overall identification ability of the model and takes values from 0–1: 1 is the best predictive model; 0.5 is a stochastic predictive model; and 0 is a not applicable model. IBS is used to evaluate the accuracy of the model, in which the numerical value IBS <0.25 represents the useful model, and 0 is the best possible value. The SHF extracted by deep learning methods have better recognition ability and accuracy in the CoxPH model compared to the texture features extracted by traditional methods.

Table 3.

Comparison of results between different methods.

Model		C-index	IBS
CoxPH	Traditional method Deep learning method	0.611 (0.601–0.618)	0.073 (0.069–0.078)
CoxPH	Traditional method Deep learning method	0.884 (0.880–0.889)	0.025 (0.020–0.031)
N-MTLR	Traditional method Deep learning method	0.630 (0.621–0.638)	0.065 (0.060–0.069)
N-MTLR	Traditional method Deep learning method	0.861 (0.840–0.879)	0.024 (0.021–0.028)

Values in parentheses are 95% confidence intervals.

IBS, Integrated Brier Score.

The two different RAD score files obtained by the deep learning method and the traditional method were put into the software X-tile version 3.6.1 (https://medicine.yale.edu/lab/rim-m/research/software) to get the cutoff value; thus the population can be divided into high- and low-risk groups. The specific numbers are shown in Fig. 3. According to the characteristics of the high- and low-risk population and corresponding labels, the KM curve is shown in Fig. 4 (left: deep learning method, right: traditional definition method).

Fig. 3.

Classification results of high- and low-risk populations based on characteristics obtained by different methods.

Fig. 4.

The KM curve and the effect diagram. After selecting useful predictive features extracted from the crowd by Lasso regression, the KM curve was drawn and the effect diagram obtained by two different methods. (Left) Deep learning method; (right) traditional method.

Further verifying the stability of SHF

To further verify the stability of SHF, we put SHF by the deep learning method and texture features obtained by traditional method into N-MTLR, respectively. In Table 3, the traditional method was based on texture features, and the deep learning method was based on the SHF. The specific results of the validation data on the N-MTLR model are shown in Table 3. The results are as follows: SHF versus texture features: C-index: 0.861 versus 0.630; IBS: 0.024 versus 0.065).

The characteristics obtained by the traditional method (texture features) and deep learning method (SHF) predict the survival probability of the same person. The abscissa is time (unit month) and the ordinate is survival probability. The blue line represents the curve obtained by the deep learning method, the red line represents the curve obtained by the traditional method, and the black dotted line represents the occurrence time of the event of interest. The results are shown in Fig. 5.

Fig. 5.

Comparison of predicting survival probability by traditional methodology (texture features) and deep learning methodology (SHF) using N-MTLR. T is the occurrence time of the event of interest, and the feature extracted by deep learning method is more consistent with the actual probability prediction. SHF, self-learning high-throughput features.

Decision curve analysis

We constructed four diagnostic models, clinical, textural features, SHF, and SHF+ clinical. The decision curve analysis (Fig. 6) shows that the clinical data model and the traditional texture feature have a similar total net income, the traditional texture feature is slightly better than the clinical data model, and the total net income of the SHF extracted by the deep learning method is the highest.

Fig. 6.

The decision curve analysis of the SHF, texture features, clinical, and SHF combined clinical model. The total net income of SHF is higher than texture features. SHF, self-learning high-throughput features.

Discussion

In this study, we sought to evaluate the prognostic value of preoperative CT image texture features and SHF on postoperative OS in the treatment of patients with CRC, thus providing a reference for doctors to make further decisions. The results show that SHF extracted from CRC preoperative CT can be used as a radiomics signature to predict OS and thus guide the prognosis. SHF outperformed the texture features both in terms of overall recognition ability and model accuracy. The DCA curve proves that SHF has a higher income.

Survivability prediction, one of the three main tasks of cancer prognosis, predicts the outcomes (survival or death) of patients with cancer (20). The objective of an accurate prognosis is to provide the patient and their family with information about the patient, including care priorities and expectations (21). Second, it is beneficial to assist with clinical decision-making (22). Further, it is useful when comparing outcomes between similar patients and determining whether patients are eligible to participate in care programs (23). Finally, it is helpful for the design and analysis of clinical trials, as well as the formulation of relevant policies. CRC is a heterogeneous and molecularly complex disease (24). CT colonography is a minimally invasive imaging examination, which can be used for the screening and diagnosis of CRC (25). The CT findings of CRC with different histological grades were significantly different. It is proposed that such profound analyses and mining of image feature data will reveal quantitative predictive or prognostic associations between images and medical outcomes (26). The analysis of texture features refers to the use of various mathematical methods to evaluate the gray intensity and position of pixels in the image so as to provide the measurement of intra-lesion heterogeneity (27). It seems robust and not significantly impacted by the technical parameters of CT scanning in some prior studies (28). In recent years, deep learning has been gaining increasing attention in the area of medical imaging because of the promising performance for big data and image processing (29). It has turned out to be very good at discovering intricate structures in high-dimensional data (13). Herein, a feature self-learning network based on multi-task training mechanism is constructed to obtain SHF. By sharing representation among related tasks, the model can better summarize the original tasks (30). Compared with texture features, SHF has a better effect. Our work achieves a more accurate prediction of OS for patients with CRC, making important clinical implications.

Our data sample consisted of CT image information and corresponding clinical information for 810 patients. The CT images were taken by GE and Siemens machines. Therefore, our approach is not limited to different scanning devices. We included all information in a complete and accurate follow-up record. This is a strength and a limitation of the experiment. This improves the accuracy of exploring the impact of relevant clinical factors on survival, but it imposes some limitations on the completeness of our information. Our proof-of-concept study is also limited by the size of the dataset, which is relatively small for deep learning. To better validate the value of our research, future goals are to re-evaluate the project by increasing the sample size and to introduce external validation sets for testing. In addition, some CT slices in our datasets were taken in 2011. Because time is also a condition that can affect accuracy, our further experiments need to be validated on datasets with a more recent time range.

In conclusion, compared with texture features, the self-learning high-throughput features (SHF) as an independent signature gives a better performance for predicting the postoperative OS in patients with CRC. SHF can be used as an independent potential signature to guide clinical practice.

Footnotes

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) received no financial support for the research, authorship, and/or publication of this article.

ORCID iD

Xiang Pan

References

Bray

Ferlay

Soerjomataram

, et al. Global cancer statistics 2018: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin 2018;68:394–424.

Wang

, et al. A tree ensemble-based two stage model for advanced-stage colorectal cancer survival prediction. Inf Sci (Ny) 2019;474:106–124.

Francis

Harold

. A new initiative on precision medicine. N Engl J Med 2015;372:793–795.

Halabi

Lin

Kelly

, et al. Updated prognostic model for predicting overall survival in first-line chemotherapy for patients with metastatic castration-resistant prostate cancer. J Clin Oncol 2014;32:671–677.

Benson

Venook

Al-Hawary

, et al. Rectal cancer, version 2.2018, NCCN clinical practice guidelines in oncology. J Natl Compr Cancer Network 2018;16:874–901.

Eccles

Welch

. Metastasis: recent discoveries and novel treatment strategies. Lancet 2007;369:1742–1757.

Yanqi

Zaiyi

Lan

, et al. Radiomics signature: a potential biomarker for the prediction of disease-free survival in early-stage (I or II) non—small cell lung cancer. Radiology 2016;281:947–957.

Oakden-Rayner

Gustavo

Taryn

, et al. Precision radiology: predicting longevity using feature engineering and deep learning methods in a radiomics framework. Sci Rep 2017;7:1648.

Lambin

Rios-Velazquez

Leijenaar

, et al. Radiomics: extracting more information from medical images using advanced feature analysis. Eur J Cancer 2012;48:441–446.

10.

Altazi

Fernandez

Zhang

, et al. SU-E-J-258: prediction of cervical cancer treatment response using radiomics features based on F18-FDG up-take in PET images. Med Phys 2015;42:3326.

11.

Aerts

Velazquez

Leijenaar

, et al. Decoding tumour phenotype by noninvasive imaging using a quantitative radiomics approach. Nat Commun 2014;5:4006.

12.

Ito

Ohki

Hayashi

, et al. Trabecular texture analysis of CT images in the relationship with spinal fracture. Radiology 1995;194:55–59.

13.

LeCun

Bengio

Hinton

. Deep learning. Nature 2015;521:436–444.

14.

Yushkevich

Gerig

. ITK-SNAP: an interactive medical image segmentation tool to meet the need for expert-guided segmentation of complex medical images. IEEE Pulse 2017;8:54–57.

15.

Tibshirani

. The lasso method for variable selection in the Cox model. Stat Med 1997;16:385–395.

16.

Fotso

. Deep neural networks for survival analysis based on a multi-task framework. arXiv 2018. Available at: https://arxiv.org/abs/1801.05512.

17.

Mark

Benjamin

Roger

. Decision curve analysis. JAMA 2015;313:409–410.

18.

Camp

Dolled-Filhart

Rimm

. X-tile: a new bio-informatics tool for bio-marker assessment and outcome-based cut-point optimization. Clin Cancer Res 2004;10:7252–7259.

19.

Dessai

Simha

Patil

. Stepwise cox regression analysis in SPSS. Cancer Res Stat Treat 2018;1:167–170.

20.

Kourou

Exarchos

, et al. Machine learning applications in cancer prognosis and prediction. Comput Struct Biotechnol J 2015;13:8–17.

21.

Glare

Sinclair

Downing

, et al. Predicting survival in patients with advanced disease. Eur J Cancer 2008;44:1146–1156.

22.

Zahuranec

Brown

Lisabeth

, et al. Early care limitations independently predict mortality after intracerebral hemorrhage. Neurology 2007;68:1651–1657.

23.

Hardy

, T urner

Saunders

, et al. Prediction of survival in a hospital-based continuing care unit. Eur J Cancer 1994;30:284–288.

24.

Punt

Koopman

Vermeulen

. From tumour heterogeneity to advances in precision treatment of colorectal cancer. Nat Rev Clin Oncol 2017;14:235–246.

25.

Pickhardt

Hassan

Halligan

, et al. Colorectal cancer: CT colonography and colonoscopy for detection--systematic review and meta-analysis. Radiology 2011;259:393–405.

26.

Virendra

Yuhua

Satrajit

, et al. Radiomics: the process and the challenges. Magn Reson Imaging 2012;30:1234–1248.

27.

Davnall

Yip

Ljungqvist

, et al.

Assessment of tumor heterogeneity: an emerging imaging tool for clinical practice?

Insights Imaging 2012;3:573–589.

28.

Miles

Ganeshan

Griffiths

, et al. Colorectal cancer: texture analysis of portal phase hepatic CT images as a potential marker of survival. Radiology 2009;250:444–452.

29.

Anthimopoulos

Christodoulidis

Ebner

, et al. Lung pattern classification for interstitial lung diseases using a deep convolutional neural network. IEEE Trans Med Imaging 2016;35:1207–1216.

30.

Ruder

. An overview of multi-task learning in deep neural networks. arXiv 2017. Available at: https://arxiv.org/abs/1706.05098.