The Promise and Future of Radiomics for Personalized Radiotherapy Dosing and Adaptation

Abstract

Quantitative image analysis, also known as radiomics, is a field that aims to analyze large-scale quantitative features extracted from acquired medical images using hand-crafted or machine-engineered feature extraction approaches. Radiomics has great potential for a variety of clinical applications in radiation oncology, which is an image-rich treatment modality that utilizes computed tomography (CT), magnetic resonance imaging (MRI), and positron emission tomography (PET) for treatment planning, dose calculation, and image guidance. A main promising application of radiomics is in predicting treatment outcomes after radiotherapy such as local control and treatment-related toxicity using features extracted from pre-treatment and possibly during-treatment images. Based on these individualized predictions of treatment outcomes, radiotherapy dose can be sculpted to meet the specific needs and preferences of each patient. Additionally, radiomics can aid in tumor characterization and diagnosis for personalized targeting, especially for identifying targets or high-risk regions within a tumor (habitats) that cannot be easily discerned based on size or intensity alone. Radiomics based treatment response prediction and assessment can also aid in developing personalized adaptated fractionation and recommending optimal dose adjustments. However, to make radiomics models more applicable across different institutions with varying scanners and patient populations, it would require further efforts to minimize uncertainties within the imaging data through harmonization and standardization of acquisition protocols.

Introduction

Quantitative image analysis, also known as radiomics, channeling its ‘omics’ counter parts from molecular biology (genomics, proteomics, metabolomics, etc.) aims to interrogate acquired image scans during diagnostic or therapeutic procedures for reusable datapoints (1). The large-scale extraction of these datapoints from known regions of interest (ROI) can be achieved using hand-crafted or machine-engineered feature extraction(2). The hand-crafted feature extraction is based on intensity, shape, size (volume), and texture describing the geometric properties and the distribution of intensities of the ROI in relation to their spatial distribution. Common examples of these features include first-order voxel intensity metrics (e.g., mean, minimum, skewness, etc.) or second-order features focusing on the statistical interrelationships between neighboring voxels (texture patterns) within the ROIs (e.g., gray level co-occurrence matrix (GLCM), gray level run length matrix (GLRLM), grey level size-zone matrix (GLSZM), and neighborhood gray tone difference matrix (NGTDM), etc.). The adopted definition of these features and their nomenclature follows the Image Biomarker Standardization Initiative (IBSI) (3). The machine engineered features are typically extracted using deep learning techniques using convolutional neural networks (CNNs) and their variants, where such features are automatically derived as part of the overall machine learning task.

The radiation oncology field is an image-rich treatment modality, which makes a fertile environment for application of radiomics as discussed later in this article. Computed tomography (CT) is the most commonly used modality for treatment planning, dose calculation, image guidance as well as onboard setup correction. These CT images could be fan beam (used for simulation) or cone-beam onboard imaging (used for verification). Other modalities are also incorporated into the radiotherapy workflow such magnetic resonance (MRI) or positron emission tomography (PET) initially for improved target definition during treatment planning and more recently for image-guidance during delivery (MR- or PET-Linac). Moreover, 4D-imaging with amplitude or phase binning are also being used in radiotherapy to account for organ movements, especially due to respiratory motion. Applications of radiomics in radiotherapy have varied from treatment planning to response prediction, with more focus on the latter as discussed here.

Radiomics for Personalized Dosing

Radiomics has been investigated as a means to predict treatment outcomes in patients undergoing radiation therapy. By leveraging radiomic features, the predicted outcome can be used to customize and personalize the prescribed radiation dose based on the risk of treatment failure. In cases where patients are at a high risk of local failure, the prescribed dose can be increased to decrease the likelihood of local recurrence. On the other hand, if the tumor is sensitive to radiation and has a low risk of local recurrence, the prescribed dose may be reduced to minimize treatment-related toxicity.

One such study explicitly explored the potential use of image-based features for individualized radiotherapy dose in lung cancer patients treated with stereotactic body radiation therapy (SBRT) (4). Specifically, this study aims to estimate a dose termed as iGray (in units of biological equivalent dose) that results in a probability of failure below 5% at 24 months for lung cancer patients after SBRT based on pre-treatment CT imaging features learned by a multitask deep neural network and clinical features. Instead of using radiomics features extracted from tumors to directly build a model for local failure prediction, conventional radiomics features were used as the label to train a multi-task deep neural network for two tasks: 1) using imaging features extracted by a CNN-based encoder to predict local failure, and 2) enforcing estimated radiomics features obtained by a decoder from the imaging features learned by CNN in the latent space to be equivalent to the handcrafted radiomics features calculated from regions of interests (e.g., gross tumor volume). An imaging-based risk score calculated form learned features by the multi-task deep neural network (termed as Deep Profiler) was then combined with biological equivalent dose (BED) to modulate the risk of failure in a multivariable regression model, where iGray was defined as the dose that results in a probability of failure below 5% at 24 months. Using 849 patients as the training cohort the model using Deep Profiler and clinical variables achieved a concordance index (C-index) of 0.72 for predicted local failures after SBRT, outperforming models using classical radiomics or clinical variables alone. Using 95 patients as the validation cohort, Deep Profiler achieved C-index of 0.77 for predicting treatment failures. Estimated iGrays indicated that dose reductions could be feasible in 23.3% of patients.

While not explicitly modeling dose-response relationships, many radiomics-based outcome prediction models can be potentially used for exploring personalized radiotherapy dosing. For instance, for patients at high risk for local failure treated with a population-based standard dose, the prescribed dose could be increased such that the probability of local failure could be reduced. Radiomics-based local failure prediction after definitive radiation or chemoradiation therapy has been explored for many cancer types. In head and neck squamous cell carcinoma (HNSCC), a radiomics model using features extracted from pre-treatment PET and CT as well as clinical variables achieved an area under the receiver operating characteristic curve (AUC) of 0.69 for local regional recurrence prediction after definitive radiation or chemoradiation therapy based on a random forest classifier (5). Using the same dataset as in reference (5), a multi-classifier based radiomics model achieved an AUC of 0.77 for local regional recurrence prediction where the model combined the output from support vector machine, discriminant analysis, and logistic regression through evidential reasoning (6). Meanwhile, incorporating during-treatment imaging such as PET (7) and CBCT (8) has also shown benefits in identifying patients at high-risk for local regional failure. In cervical cancer patients, a radiomics model utilizing features extracted from pre-treatment PET and multi-parametric MRI achieved a high accuracy (>0.95 AUC) on two external cohorts for prediction of locoregional control in locally advanced cervical cancer after definitive chemoradiation therapy (9).

In addition to local control, personalized radiotherapy dosing must also take into account the potential for toxicity of normal tissues/organs. Based on predicted toxicity after radiotherapy, the treatment plan can be adjusted and optimized to reduce the likelihood of radiation induced toxicity. Radiomics based toxicity prediction after radiotherapy has been explored for a number of different disease sites (10,11). Generally, characterizations of spatial radiation dose distribution within an organ (e.g., learned by a neural network) often outperform dose-volume based parameters in toxicity prediction (12). Radiomics features extracted from pre-treatment images have shown promise in improving toxicity prediction beyond the analysis of dose distribution alone. A recent review article has summarized the applications of radiomics in predicting toxicity related to parotid glands in head and neck radiotherapy, radiation induced lung injury in thoracic radiotherapy, heart disease in breast radiotherapy, and radiation-induced lower gastro-intestinal toxicity in prostate radiotherapy (11).

Ultimately, in order to develop personalized treatment strategies, it is important to establish a dose response function and normal tissue complication probability based on the unique features of individual patients, rather than relying solely on population-based dose response curves. This personalized approach should take into account not only radiomics features, but also other relevant patient characteristics. By providing individualized predictions of tumor control and toxicity probability, clinicians can tailor treatment plans to meet the specific needs and preferences of each patient.

Radiomics for Personalized Targeting

Imaging plays a critical role in defining targets for radiation therapy, and radiomics has emerged as a promising tool for this purpose. Radiomics-based analyses are particularly valuable for identifying targets that cannot be easily discerned based on size or intensity alone on imaging.

For instance, when identifying treatment targets in lymph nodes, some nodes may be clearly malignant due to their size or high uptake on PET, but smaller or less avid nodes on PET may be more difficult to classify. Partially due to the uncertainties in identifying malignant lymph nodes that truly harbor cancer cells, elective neck irradiation is still performed to treat cervical nodal basins that appear normal in radiotherapy for HNSCC, leading to substantial treatment-related toxicity. Many attempts have been made to reduce target volumes in treating HNSCC. A recent study (13) called INRT-AIR (Involved Nodal Radiotherapy using AI-based Radiomics) aims at treating malignant nodes only by avoiding elective neck irradiation to spare a large portion of normal neck tissues. An involved nodal radiation therapy treatment strategy requires a reliable approach to differentiate malignant nodes from benign ones. In the INRT-AIR study, a hybrid model (14) utilizing both handcrafted radiomics features and learned features by CNN from PET and contrast-enhanced CT was developed to identify malignant nodes among those either of small size or less fluorodeoxyglucose (FDG)-avid on PET. The model was trained on 791 lymph nodes from 129 patients that underwent neck dissection where the lymph node malignancy statuses were determined from the pathology reports (15). Radiomics- and CNN-based models were trained separately, and the final output probabilities of the testing samples were calculated through a fusion of the output probabilities from the radiomics and CNN models with weighting factors determined from the model performance measured by AUC on the validation set. Under the guidance of this hybrid lymph node malignancy prediction model, the INRT-AIR enrolled 67 patients.

An example in Figure 1 illustrates the dosimetric advantages of INRT-AIR for HNSCC as compared to conventional treatment, which includes elective neck irradiation. Preliminary results of this trial have been reported at the 63^rd annual meeting of the American Society for Radiation Oncology (16). At a median follow-up of 12.8 months, there have been no solitary elective lymph node recurrences (16). Acute toxicity was modest, with only 10% of patients reporting grade 2 dermatitis, and only 21% of patients requiring a gastrostomy tube, with median removal after 2.8 months for disease-free patients. From a quality-of-life perspective, the mean composite MD Anderson Dysphagia Inventory (MDADI) scores at 12 months are much higher than a cohort of patients treated with standard IMRT with elective neck irradiation from a prospective cohort at Royal Marsden (17). These results suggest that involved nodal radiation therapy enabled by radiomics for personalized nodal structure targeting offers a highly effective method to improve a crucial quality-of-life outcome in HNSCC.

Figure 1.

Dosimetric comparison for a patient treated by INRT-AIR (a and b) with personalized targeting or conventional treatment that includes elective neck irradiation (ENI) (c and d). (a) and (b) display the actual plan used to treat the patient with INRT. Note that the numbers in the figure are in units of cGy.

Beyond lesion and node malignancy classification for target identification, radiomics has also been explored to characterize tumor heterogeneity by classifying the whole tumor to different zones of different risk levels. Such a classification scheme could be used for dose painting, where a higher radiation dose can be delivered to regions at high risk of recurrence. To identify treatment resistance regions within the primary target in HNSCC, Bogowicz et al. divided the whole tumor in pre-treatment contrast-enhanced CT to 8 sub-volumes and used sub-volume radiomics features (intensity and texture) to differentiate radioresistant and controlled regions, achieving 0.70 AUC in the validation cohort (18). In another example (19), using imaging features extracted from multi-parametric MRI (including both DCE-MRI and diffusion weighted imaging), Stoyanova et al. developed a prostate tumor habitat risk scoring system to predict the likelihood of cancer of individual pixels, which provides a valuable tool for defining radiation therapy boost volumes in prostate cancer patients.

Radiomics for Adaptive Radiotherapy

Adaptive radiotherapy (ART) has been an advancing topic since the late 1990s. Yan et al. presented the concept of ART to improve radiation treatment by systematically monitoring treatment variations and incorporating them to re-optimize the treatment plan early on during the course of treatment (20). This intuitively falls into personalized radiotherapy since the treatment strategy will be modified based on new information gathered for the individual patient during treatment. The target and organs at risk for radiotherapy are dynamic. It varies in position, shape, size, and biology over a time frame that also varies (21). New information learned during the treatment includes anatomical changes (e.g., motions) and deformations of the organs and biological changes (e.g., changes in prostate specific antigen for prostate cancer). Thus, there are mainly two types of ARTs, adaptation based on anatomy or response (biology). Radiomics has been explored for both anatomy-based ART and response-based ART.

Anatomy-based Adaptation

ART is a labor-intensive process with segmentation needed for adaptation, which is often manual. There is also a tradeoff of offline or online adaptation as online adaptation can potentially provide greater treatment precision because it uses real time images during treatment. However, it will increase the daily effort and treatment time. As pointed out by Schwartz et al, offline ART appears to be more practical for disease that the anatomy changes slowly, such as head and neck radiation treatment (22).

Due to the resource-intensive nature of ART and as not every patient may benefit from ART, radiomics has been explored to identify those patients that could truly benefit from ART. One study was implemented to determine tumor biomarkers using radiomics from pre-treatment MR images for predicting ART eligibility in nasopharyngeal carcinoma (23). The same group further investigated multi-organ and multi-omics analysis for identification of ART eligible patients (24). They extracted multi-omics data (radiomics, morphology, dosiomics, and contouromics) from eight organs. The radiomics model was found to play a dominant role for ART eligibility in nasopharyngeal carcinoma patients. Another study also utilized radiomics analysis to distinguish patients likely or unlikely to demonstrate early radiation-induced tumor regression before treatment so that frequent online ART could be scheduled only for early “regressors” (25). They used CNNs with GTV input pretrained on natural images to extract deep features, then these features were trained by machine learning models to predict regression versus nonregression. The workflow is shown in Figure 2.

Figure 2.

Workflow of deep learning radiomics approach for adaptive treatment from (25). Axial CT slices of primary gross tumor volume (GTVp) and nodal gross tumor volume (GTVn) were inputted to 16 convolutional neural network (CNN) deep learning models pretrained on natural images, and deep features were extracted. Next, the most robust features were selected and redundant features eliminated (selection step 1). Then, the top 10 features for each of the five feature selection algorithms were identified (selection step 2). Machine learning algorithms were used to predict primary GTV (GTVp) and nodal GTV (GTVn) regression versus nonregression.

In addition to patient triage, studies have been carried out to enable early detection of disease response so that the most effective adaptive treatment can be delivered. For instance, one study focused on early detection of treatment response for pancreatic cancer patients by combining delta radiomics from daily CT images with cancer antigen 19-9 biomarkers to allow more time for necessary plan adaptation (26). Zhang et al. also showed a longitudinal CT radiomics model could model the trend of tumor change in non-small cell lung cancer (NSCLC) and could be potentially used for treatment decision with adaptation (27). Another study extracted radiomics features from longitudinal MR scans acquired during MR guided radiation therapy and showed the potential predictive power of delta radiomics features for the treatment response in pancreatic cancer (28).

Biology-based Adaptation

Although biology-based adaptation has been discussed for a long time, it remains challenging to effectively incorporate the ever-increasing amount of biological and response information into radiation treatment planning and adaptation. However, this is an essential component for personalizing treatment strategies, in addition to geometry-based adaptation. Radiomics has the potential to add useful information to other biology biomarkers. Tseng et al. presented a methodology for knowledge-based response-adaptive radiotherapy (KBR-ART). Using dosimetric data, radiomics, and biological data as input, linear or nonlinear feedback models and stochastic or reinforcement learning (RL) methods can be applied for treatment adaptation in radiation therapy (29). The same group developed a deep reinforcement learning model to do automate dose adaptation for NSCLC using clinical, genetic, dosimetric, and imaging radiomics features (30). This approach was benchmarked against real clinical decisions that were applied in an adaptive dose escalation clinical protocol and achieved similar results to the benchmark. Luo et al. used a multi-objective Bayesian network (MO-BN) to identify important features such as PET radiomics, dose, cytokines, etc. for local control and toxicity joint prediction for NSCLC to aid decision support systems for personalized response-adapted radiation therapy (31). Niraula et al. developed a quantum deep reinforcement learning framework for clinical decision support that can estimate an individual patient’s dose response mid-treatment based on multi-omics data and recommend an optimal dose adjustment (32).

Functional and molecular imaging techniques offer the ability to obtain biological information non-invasively before, during and after treatment, which can be utilized for treatment adaptation. Radiomics features extracted from these images may provide better biologically relevant and targeted information to aid in treatment adaptation. Forouzannezhad et al. used longitudinal radiomics data from FDG-PET, CT, and perfusion single photon emission computerized tomography (SPECT) images to predict survival outcome for supporting adaptive treatment in NSCLC patients (33). Mierzwa et al. conducted a randomized phase II multicenter clinical trial to test the hypothesis that physiologic MRI-based radiotherapy dose escalation would improve the outcome of head and neck cancer patients with poor prognosis (34). Their findings showed that a physiologic MRI-based radiation therapy boost decreased local regional failure without a significant increase in grade 3+ toxicity or longitudinal patient reported outcome differences. Another phase 2 clinical trial adapted the treatment plan to escalate dose to the FDG-avid tumor detected by midtreatment PET, which resulted in favorable local-regional tumor control (35).

Uncertainties to Minimize for Large-scale and Multi-institutional Radiomics Models

As radiomics continues to expand and show promise, it is important to minimize uncertainties within the data. This allows models to be more applicable across different institutions with different scanners and different patient populations. It also ensures that a true signal is being captured in the model and not simply noise due to the differences in scanning protocols.

Image Acquisition and Reconstruction

As patients undergoing radiotherapy typically have a simulation CT for treatment planning, the majority of radiomics studies in radiotherapy applications often involve CT images. Therefore, this section will focus on image protocol impacts on radiomics for CT. For MRI and PET, other reviews on these radiomics uncertainties can be consulted (36). CT image acquisition and reconstruction settings include tube voltage (kVp), tube current (mAs), reconstruction kernel, and voxel size, all of which are part of the image protocol. Mackin et al. created a specially designed radiomics phantom (Credence Cartridge Radiomics (CCR)) comprised of 10 different cartridges each with different materials in order to produce a wide range of radiomics feature values (37). This radiomics phantom was scanned on 17 different CT scanners and radiomics values were compared to features from NSCLC tumors from 20 patients. The variability in the radiomics features from different CT scanners was found to be comparable to the variability in features from NSCLC tumors for some features, thus indicating the need to incorporate interscanner differences and minimize these differences.

To investigate the impact of tube current and tube voltage, Fave et al. simulated changes in tube current and tube voltage for 20 NSCLC patients (38). The tube voltage effect was adjusted using an offset value for different tube voltages based on predicted CT intensity values of tumors for different tube voltages based on the atomic composition. It was found that tube voltage would likely not impact features. Tube current was modeled using Gaussian noise and was found to statistically significantly impact features. Larue et al. used the CCR phantom on 9 different CT scanners with varying tube current and did not find feature values to be affected by tube current (39). Further, they found that optimization of gray-level discretization could be performed to potentially improve the prognostic value of radiomics features without compromising feature stability. In further analysis with the CCR phantom, Mackin et al. found that tube current had more effect on features from homogenous materials, such as acrylic, than on materials that better matched the texture of tumors (40). Therefore, overall, tube current and tube voltage are unlikely to impact radiomics features of patient tumors or organs at risk as long as there is enough variability in CT values to have more texture than homogenous materials.

Zhao et al. evaluated the impact of reconstruction kernels on lesions within an anthropomorphic phantom and later repeated this analysis in 32 lung cancer patients (41,42). Both phantom and patient studies found that sharp and smooth kernels should not be used together in radiomics studies as the different reconstruction kernels can significantly affect features. Lu et al. also used this patient dataset and found the highest agreement (concordance correlation coefficient (CCC) > 0.8) between groups using the same reconstruction kernel while the lowest agreement (CCC < 0.51) was when the reconstruction kernel and voxel size was varied (43). Voxel size has also been demonstrated to be an issue and impact radiomics features in studies utilizing the CCR phantom (39,44). Both studies also showed that resampling the image prior to radiomics feature extraction could remove this feature dependence on voxel size. However, Mackin et al. demonstrated that resampling alone was not sufficient, images should be resampled and filtered before feature extraction (45). Figure 3 demonstrates the variability from different scanners using no resampling and no filter, just resampling, and resampling with Butterworth smoothing. With no correction the average scaled variability was 0.51, with resampling the average scaled variability was 0.96, while with resampling and filtering the average scaled variability was 0.2. This was also shown to impact clustering of 8 NSCLC patients each with 5 field of views. Only smoothing and filtering was able to correctly cluster the 40 CT images from the patients.

Figure 3.

Figure from (45). Scaled contrast for the CCR phantom’s rubber particle cartridge scanned with 17 different CT scanner configurations. (a) Feature values without image preprocessing. (b) Feature values calculated after all images had been resampled to 1 mm/pixel. (c) Feature values calculated after all images had been resampled to 1 mm/pixel and filtered with Butterworth filter (order 2, frequency cutoff 75). The points are color coded and labeled according to the manufacturer of the scanner: GE indicates GE Healthcare (green); P, Philips Healthcare (purple); S, Siemens Healthineers (pink); T, Toshiba Medical Systems (cyan).

Ger et al. used an updated version of the CCR phantom that had 6 cartridges of various material encased in a high-density polystyrene buildup with the dimensions to match the mean European woman’s chest (46). As Mackin et al. demonstrated the variability of lung protocols on different scanners (37), the updated phantom was utilized to determine if a controlled protocol could reduce variability. Incorporating all of the above studies, a controlled protocol was established (Table 1) and the phantom scanned on 100 CT scanners. This protocol in addition to resampling was able to reduce the overall variability by more than 50% compared to the local lung or head protocols for each scanner. Conversely, Orlhac et al. have published a methodology called ComBat which compensates for the variability in protocols and have published the methodology and validation of it for CT, MRI, and PET (47-49). There are several assumptions that must be met for ComBat to work, such as that the distributions of features are similar except for a shift and spread and that different features have to be independent.

Table 1.

Settings for Controlled Radiomics Protocol

Manufacturer	kVp	mAs	Scan Type	Spiral Pitch Factor	Approximate CTDIvol (mGy)	Convolution Kernel	Display Field of View (cm)	Slice Thickness (cm)
GE	120	200	Helical	1	13.3	Standard	50	2.5
Philips						C		3
Siemens						B31s, B31f
Toshiba						FC08, FC18

Patient artifacts can also be a concern for radiomics studies, such as streak artifacts from implants. While metal artifact reduction reconstruction algorithms can improve these greatly, they may not completely remove them. This can particularly impact images of the head where dental artifacts cause streaks through targets for head and neck cancer patients. Inclusion of slices with artifacts can lead to noise within models and less precise models as the actual texture is not visible. However, as these artifacts can impact many patients, excluding all patients with artifacts may dramatically reduce the patient population available for modeling and the later applicability of the model to the general clinic population. It has been shown that up to 50% of the original target slices can be removed without impacting radiomics features (50). In over 450 head and neck patients, following this guidance was found to only remove 3% of patients, thus this methodology can be helpful for keeping potential patients for radiomics modeling high while not introducing extraneous noise into the model from artifacts.

Radiomics Software Variability and Data Availability

There are many different radiomics software available with many research groups using their in-house solutions. Foy et al. demonstrated that there were significant differences when comparing features among two in-house solutions at the University of Chicago and two freely available radiomics packages (MaZda and IBEX) (51). It was also found in this study that the package defaults for feature categories, such as GLCM, can be vastly different. Therefore, care must be taken when utilizing software to ensure settings are appropriately chosen for the given analysis. In a follow-up analysis, 105 esophageal cancer patients were analyzed with logistic regression to classify radiation pneumonitis with radiomics features being extracted from one in-house software and two freely available radiomics packages (IBEX and PyRadiomics) (52). It was shown that there were differences in classification ability of the features extracted from the different software.

These differences in features and therefore performance of the different software, led to the creation of IBSI (3). IBSI used 25 research teams with different radiomics software. Iteratively, the features were analyzed from a digital phantom and CT scan of a lung cancer patient. Initially, the agreement from the different software was poor with 76.8% of the features from the digital phantom and 65.4% of the features from the lung cancer patient having weak consensus. After iteratively adjusting features within each of the software, the weak consensus was only 0.4% of features from the digital phantom and 1.4% of the features from the lung cancer patient image. IBSI provides definitions for each feature and reference values for features for from these images and a publicly available dataset of 51 patients for researchers to compare their radiomics software against. By completing IBSI compliance, it makes it easier to draw conclusions from different radiomics studies as the features then report the same information.

The 51 patients that IBSI references are from The Cancer Imaging Archive (TCIA). As radiomics studies continue to expand, it is important to incorporate data from different institutions to allow the most applicability of the developed models. TCIA is sponsored by the National Cancer Institute (NCI) and thus serves as a valuable resource as it contains both patient and phantom images related to cancer. The Imaging Data Commons is also sponsored by the NCI and could be a good resource for finding additional images. There are many other sources not sponsored by a government agency that can be used to find additional data, such as previous Medical Image Computing and Computer Assisted Intervention (MICCAI) challenges and figshare.

Future of Radiomics Features

Most radiomics studies to date have utilized conventional radiomics features. These include first order histogram features (e.g., energy, kurtosis), GLCM (53,54), GLRLM (55), and NGTDM features (56). Many of these conventional features were designed for images that are not medical images. The Haralick features and Galloway features which compose the GLCM and GLRLM features were designed for aerial or satellite images. The NGTDM features were also used for classifying land within images. Therefore, almost all traditional radiomics features used are optimized for use in other image types.

While there has been success found using these features on medical images, many of these studies have not reached clinical significance levels to be utilized clinically. A main drawback to these features is their tenuous relation to any meaning in a medical sense. Recently, deep learning has been used for “radiomics” studies. The features being highlighted in these deep learning models do not have the same drawbacks as traditional radiomics features. However, many deep learning models are black boxes which again do not necessarily have a clinically meaningful feature that is being selected or a way to understand why such a feature may be selected. Deep learning models that allow understanding of the filters utilized in each step to understand what features are being extracted and what this may mean clinically are therefore needed. Another avenue is feature development to continue traditional radiomics studies, but with biologically relevant designed features. Incorporating these types of features in combination with the image acquisition and reconstruction techniques described above to reduce noise within model building and analysis on multi-institutional data will allow radiomics to reach its next stage and allow clinical implementation of radiomics-based tools across clinics.

Discussion

Radiomics is a powerful tool that can enable personalization of radiotherapy treatment in terms of personalizing patient dose prescription, improving cancer targeting, as well as adapting dose fractionation to changes in anatomy during a radiotherapy course. Most of the existing literature has focused on hand-crafted approaches, in which known computer vision features are derived from single or multi-modality images (57). However, the newly emerging literature in radiomics is gearing towards more machine-engineered approaches with machine learning techniques or a hybrid of both (58).

Despite the potential of radiomics to personalize radiotherapy care and improve outcomes. There are several lingering challenges that still need to be addressed as discussed in the following.

First, it is recognized that radiomics features are sensitive to image acquisition protocols and reconstruction algorithms, as discussed above. This may preclude generalization from one study to another or implementation from one institution to another. Simple procedures for standardization (or harmonization) are being presented but these may also run the risk of masking relevant features too. Hence, balancing for well-studied designs need to incorporate “robust” radiomics features. Multi-institutional studies with larger cohorts can assist in this process.

Secondly, there are also known uncertainties in the clinical endpoints that are used to build current radiomics models, which can vary from one institution practice to another, further complicating the development of robust models. Thus, more consensus on estimating clinical endpoints for radiomics is required. This may not be the only source of bias, therefore, uncertainty estimates using known statistical procedures should be included as part of the radiomics model presentation (59).

Thirdly, there is a lack of comprehensive validation studies of radiomics studies. Most of the presented literature on radiomics has focused on retrospective analyses and there is a lack of prospective studies that are needed to accelerate the process of radiomics translation into clinical practice (60). Fourthly, there are several known clinical prognostic factors that may confound the performance of radiomics studies such as tumor volume, which have impacted the generalizability of earlier studies in the field (61,62).

Last but not least, radiomics can be used in conjunction with other known clinical and other “-omics” features to optimize prediction power of such panomics (multi-omics) models (63). Radiomics remains an exciting field with tremendous possibilities for personalized radiotherapy. However, further investigations are required to unlock its full potential.