Human-Level Comparable Control Volume Mapping with A Deep Unsupervised-Learning Model for Image Guided Radiation Therapy

Abstract

Purpose:

To develop a deep unsupervised learning method with control volume (CV) mapping from patient positioning daily CT (dCT) to planning computed tomography (pCT) for precise patient positioning.

Methods:

We propose an unsupervised learning framework, which maps CVs from dCT to pCT to automatically generate the couch shifts, including translation and rotation dimensions. The network inputs are dCT, pCT and CV positions in the pCT. The output is the transformation parameter of the dCT used to setup the head and neck cancer (HNC) patients. The network is trained to maximize image similarity between the CV in the pCT and the CV in the dCT. A total of 554 CT scans from 158 HNC patients were used for the evaluation of the proposed model. At different points in time, each patient had many CT scans. Couch shifts are calculated for the testing by averaging the translation and rotation from the CVs. The ground-truth of the shifts come from bone landmarks determined by an experienced radiation oncologist.

Results:

The system positioning errors of translation and rotation are less than 0.47 mm and 0.17 degree, respectively. The random positioning errors of translation and rotation are less than 1.13 mm and 0.29 degree, respectively. The proposed method enhanced the proportion of cases registered within a preset tolerance (2.0 mm / 1.0 degree) from 66.67 percent to 90.91 percent as compared to standard registrations.

Conclusions:

We proposed a deep unsupervised learning architecture for patient positioning with inclusion of CVs mapping, which weights the CVs regions differently to mitigate any potential adverse influence of image artifacts on the registration. Our experimental results show that the proposed method achieved efficient and effective HNC patient positioning.

1. INTRODUCTION

Positioning accuracy is vital for head-and-neck cancer (HNC) radiation therapy (RT) because the anatomy is complex, and the organs-at-risks are proximate to the tumor [1, 2]. Image guidance is essential for achieving high positional accuracy and daily patient positioning CT (cone-beam CT (CBCT)) is used as pre-treatment imaging modality [3–5]. After the acquisition of the planning CT (pCT) and daily CT (dCT), image registration algorithms are employed to find the optimal transformation that maximizes various image similarity measures between these images [6]. Many researchers have proposed or evaluated different image registration methods for the positioning of HNC patients [7–10], most of which are based on global registration.

However, image similarity measures are known to be non-convex in relation to the affine transformation parameters. Thus, the convergence of the algorithm to the global optimum is not guaranteed [11]. Furthermore, many image artifacts can cause the global correspondence between pCT and dCT to be ruined. Metal artifacts, for example, will appear diverse in different CT scans depending on the scan modality (CT versus CBCT) [12], technique (kVp), and reconstruction method. In addition, the change of tissues such as tumors will shrink during the treatment. Thus, anatomical structures may not be rigorously equivalent at different time points. Inclusion of image artifacts and anatomical variations may bias global image similarity measures and compromise registration algorithms [13, 14]. The use of homologous control points on the two images to be registered is used in a variety of image registration methods [15, 16]. In fact, automatically identifying the control points on both images is complicated, time-consuming, and often error-prone due to image content changes. For these reasons, although automatic registration algorithms have long been incorporated into the image-guided radiotherapy (IGRT) system, HNC patient positioning still requires human intelligence and manual intervention for correcting positional errors [17].

Deep learning (DL) is emerging as a powerful tool for image analysis and promises to revolutionize IGRT practice [18–22]. In general, the registration parameters in traditional rigid registration methods are obtained by optimizing an objective function for a given pair of input images in an iterative manner. The DL model finds the solution through a pre-designed search trajectory by prior knowledge extracted from training data [23]. To some extent, one may think that a DL model finds the solution through a pre-designed search trajectory by prior knowledge extracted from training data [24]. Scientifically, the DL approach should be investigated as it represents a completely different avenue for achieving the goal of image registration. Because of the possibility of pre-defining the search trajectory, DL offers a few potentially useful advantages, such as higher computational efficiency, and more robust and accurate solutions. Because of elimination of iterative search, the execution of a DL model can be an order of magnitude more efficient than conventional approaches [25]. The enhancement in the robustness of solution search may arise from the fact that local optima become less a concern in DL modeling as zero gradients are often (if not always) at saddle points [18]. Of course, we are still at an early stage in using DL for image analysis and much research needs to be done to gain a comprehensive understanding of the pros and cons of the two totally different approach.

In the traditional image registration method, the dCT image and pCT image are matched without filtering or prioritizing the information in the images. In reality, the image information’s reliability is spatially heterogeneous, and some regions of the dCT image can be mapped to the pCT image more accurately than others. There are at least two types of regions in a dCT image: regions whose correspondence in the pCT image is easily identifiable, and regions mapping to the pCT image is less obvious or even impossible for a variety of practical reasons, including tissue deformation, a lack of distinct image features, and metal artifacts caused by exogenous objects such as metallic impingement. Schreibmann and Xing [26] have emphasized that these two region types should be treated differently for the pCT and dCT mapping.

The goal of introducing control volumes (CVs) in DL-based registration is similar to that in traditional registration methods, i.e., to mitigate the adverse effects of image artifacts caused by exogenous object(s) such as metallic implantation. Here we leverage the previous investigations for an improved DL model by effectively emphasizing the artifact-free control volume (CV) regions in the metric calculation [27]. We extend the CV-based image registration to build a deep learning model for automated patient positioning. The essence of the proposed data-driven approach is to pre-identify the CVs (high importance volumes) and incorporate a priori knowledge about the input data into the learning process to avoid/reduce potential registration errors induced by image artifacts. The pre-clustering of the input imaging pixels by placing CVs on the reliable region helps to mitigate the potential errors and enhance the results of the deep learning model.

This is the new attempt a DL model has been applied to generate couch shifts for mapping CVs from dCT to pCT. Previously, our team proposed and published the CV-based image registration method based on the conventional iterative method [26]. However, the serial iteration method is time-consuming. Besides, the result is easy to fall into the local minimum because the patient positioning or the rigid registration is an ill-posed problem. To alleviate the above shortcomings, in this study, we incorporated the DL method into the CV-based patient positioning. Since there is no iterative process in the test, the proposed method is efficient. The proposed method is accurate because the DL-based method has high dimensional vector space, which can alleviate the local minima.

Our works contributions are 2-fold:

• An unsupervised deep learning framework is proposed for highly accurate and efficient patient positioning in CT-guided radiotherapy.

• We incorporate a CVs mapping strategy into the DL architecture, which emphasizes the CVs regions to mitigate any potential adverse influence of image artifacts on registration.

2. METHODS AND MATERIALS

2.1. Patient data

In this study, we used a total of 158 HNC patients (554 CTs) from the Cancer Imaging Archive [28–31] and a database at Stanford Cancer Center. Each patient underwent several scans. To train and verify the network, we randomly divided all the image volumes into three sets: the training set (133 patients with 472 CTs), validation set (10 patients with 34 CTs), and testing set (15 patients with 48 CTs). The study was approved by the Stanford Institutional Review Board (#53173). Details on the datasets are provided in supplemental materials.

2.2. Patient positioning with DL model

For a given pCT, the CVs were manually placed by a treatment planner. The DL model was then carried out to map the CVs between the dCT and pCT. The final IGRT transformational parameters were calculated by averaging the translational and the rotational parameters derived from the individual CVs [26, 32].

To develop a reliable DL model for automated HNC patient positioning, four tasks, i.e., construction, training, validation, and testing of the deep neural network, need to be implemented. A DL model is constructed to predict the transformational parameters between the pCT and the dCT for patient positioning. Instead of manually annotating a large number of training images, we develop an unsupervised DL model for CV mapping. The DL model is trained by maximizing the image similarity between the pCT and dCT. The training data consists of the pCT image, dCT image, and the CVs’ location within the pCT image. The model is then evaluated on the independent dCT.

2.3. Unsupervised DL model

A schematic of the proposed unsupervised learning framework for CV mapping is shown in Figure 1(a). The three inputs are the 3D dCT and pCT images, denoting as I_d and I_p, and the location v of the CV bounding box within the pCT. The CV’s location is incorporated in the pCT image to form a new input image I_pv, which is equal to the pCT image except at voxel locations within the annotated CV bounding box where the pCT voxel values are multiplied by a constant α = 10. The dCT image I_d and annotated pCT image I_pv are used as inputs to a 3D densely connected neural network (DenseNet) N_θ with training parameters θ. The output is a spatial transformation with translation and rotation parameters φ = N_θ(I_d, I_pv) that maximize the image similarity between the transformed dCT $I_d^{\prime }$ and the pCT volumes with an emphasis on the annotated CV [33].

Figure 1.

Proposed unsupervised learning framework for CV mapping. (a) Schematic of the methodology and (b) architecture of the underlying 3D-DenseNet machine learning model.

Abbreviation: pCT = planning computed tomography, CV = control volume, L_G = the global image similarity loss, L_CV = and the CV image similarity loss.

The Dense-Net parameters θ are trained with a new compound loss function

$$L\left(\theta ;I_d,I_p,v\right)=L_G\left(\theta ;I_p,I_d^{\prime }\right)+\lambda L_{CV}\left(\theta ;I_{p,CV},I_{d,CV}^{\prime }\right),$$ (1)

which is the weighted the global image similarity loss L_G and the CV similarity loss L_CV where λ is a constant parameter that determines their relative weighting. The rest of the notation includes I_p,CV and $I_{d,CV}^{\prime }$, which are the CV images extracted from the pCT and the transformed dCT images at the same location v, respectively. Both terms in the loss function are computed using the normalized cross-correlation (NCC) metric [32] :

$$NCCloss=-\frac{\sum _p\left(I_1(p)-{\overline{I}}_1\right)\left(I_2(p)-\overline{I_2}\right)}{\sqrt{{\sum }_p{\left(I_1(p)-\overline{I_1}\right)}^2}\sqrt{{\sum }_p{\left(I_2(p)-\overline{I_2}\right)}^2}},$$ (2)

which is a normalized measure of the similarity between two images I₁ and I₂, with p the voxel index. $\overline{I_1}$ and $\overline{I_2}$ are the average value of I₁ and I₂, respectively. The optimization framework aimed to minimize the loss function. Therefore, we added a negative sign to the NCC and used it as the loss function. When the loss function is minimized, the NCC is maximized. Minimizing the global loss L_G alone, as in traditional image alignments methods, may generate large errors due to localized image artifacts or anatomical disparities that would negatively bias the global criterion. In contrast, including the local loss L_CV in the backpropagation algorithm contributes to minimization of the errors in the annotated CV that reflects a physician’s region of interest, thereby alleviating the leakage of errors from other parts of the image.

The architecture of the Dense-Net model used to estimate the optimal spatial transformation is shown in Figure 1(b). The inputs are the dCT and annotated pCT images. The model consists of a convolution layer, three dense blocks next to the transition layer and the last dense block next to the pooling layer and a linear layer. The dense blocks use dense connectivity between sequences of convolution, batch normalization, and rectified linear units. The transition layers formed by convolution layer and pooling layer are used to reduce the dimension of the feature mapping between adjacent dense blocks. The final pooling and linear layer are used to reduce the output dimension to the translation and rotation spatial transformation parameters. The readers are referred to Ref [34] for more details on the Dense-Net model and the specific architecture used in this work.

2.4. Training of the unsupervised DL model

Training of a DL model usually requires a large amount of annotated datasets, and this presents a bottleneck problem in establishing a DL model [35, 36]. In contrast to most supervised DL methods, in our approach we do not rely on annotated datasets, which significantly reduces human effort. The required datasets consist of three parts: pCT, dCT, and CVs’ position whereby pCT and dCT are acquired at a different time for a specific patient. In the training stage, the CVs’ location is randomly generated within the pCT image without manual annotation. For a given CV, the unsupervised DL model attempts to find an optimal transformation to minimize the loss function. In the testing stage, the CVs’ position is manually selected by a physician.

We train the network using an open-source machine learning framework Pytorch 0.4.1 on 4 NVIDIA GeForce GTX 1080 Ti GPUs, and 128 GB of internal memory. Training takes about 30 hours on this computational equipment. To minimize the loss function, we use Adam algorithm to obtain the optimal parameters θ [37]. The learning rate is set at 10⁻² and then gradually reduced to 10⁻⁶. A total of 10⁵ iterations are used to train the 3D Dense-net. In this work, the number of voxels of the input two-channel images is 216 × 216 × 184 × 2. All the images are resized into the voxel size of 1 × 1 × 1 mm³ (216 × 216 × 184). The physical size of the CV is 40 × 40 × 30 mm³. The batch size is 1. Random translation and rotation were introduced for data augmentation.

2.5. Testing of the unsupervised DL model

For each patient, we acquired the pCT, dCT images, and CVs, which were manually selected within the pCT. In principle, a single CV is enough to derive the translational and rotational parameters of couch shifts for rigid patient positioning. However, bony anatomies are not a simple rigid object in the HNC region. For robustness, we typically select four CVs, and the final transformational parameters are calculated by averaging the parameters of the individual CVs. Generally, the CV should be anatomically robust (e.g., contains no artifacts) and contains significant image features. A region having a CV in the brain does not contain much image texture and should thus be avoided. Another rule of thumb here is that the CVs inside the patient should be spatially distant from each other so that all anatomy regions are considered. In this study, the CVs are located in the vicinity of bony structures as follows: C2, C6, mandible, and skull.

The evaluation of patient positioning results is a challenge due to the lack of ground truth. In this study, eight known and easy to recognize anatomical points were localized in both the pCT and dCT by an experienced radiation oncologist to find a reference transformation. Each of the bony structures has two annotated landmarks. Assume that the N landmarks in the pCT and dCT are defined by ${\overrightarrow{LM}}_{pCT}(i)$ and ${\overrightarrow{LM}}_{dCT}(i)$, i = 1, 2, 3, …, N, respectively. The translational and rotational parameters using the bony landmarks alignment can be calculated:

$$min\sum _{i=1}^N{\left|R{\overrightarrow{LM}}_{dCT}(i)+T-{\overrightarrow{LM}}_{pCT}(i)\right|}^2,$$ (3)

where R indicates the rotation in the yaw, pitch, and roll directions, and T indicates the translation in anterior-posterior (A-P) and left-right (L-R), and superior-inferior (S-I) directions. To make a quantitative comparison, we use the results of the landmarks alignment as the reference. We compare the effects of the proposed population-based approach with the traditional rigid alignment process in SimpleITK, a commonly used registration toolkit, to demonstrate its performance [38]. The implementation details are shown in the Supplementary Materials section.

2.6. Statistical analysis

The box-and-whisker plot with 5^th percentile and 95^th percentile is employed to depict the statistical measurement of the conventional and the proposed approach. The overall mean (M), random error (σ), and systematic error (∑) are calculated for both the conventional and proposed approach [39, 40].

3. RESULTS

In Figure 2, we show one example of the CVs mapping results. Rows (1), (2) and (3) show the results without alignment, with conventional alignment, and with the proposed method, respectively. To visualize the alignment results, columns (a) and (b) show the checkerboard of two axial slices from images. Note that the whole cone-beam CT (CBCT) image is shifted posteriorly relative to its expected location when the conventional alignment method is used (Figs. 2(a2) and (b2)). The misalignment is evident as bony structures seen on the pCT have no correspondence on the CBCT, which is mitigated by the proposed method as shown in Figs. 2(a3) and (b3), since the DL network puts more weight on the artifact-free CVs in the loss function. Column (c) shows the superposition of the 3D rendering (differences between the images are highlighted in orange). The differences are evident in the 3D rendering of Fig. 2(c2) compared with the proposed method in Fig. 2(c3). Column (d) shows the selected yellow rectangular CVs, #1 to #4, as overlays on the background pCT image. Within each CV, the shifted CBCT image is shown. The patient structures do not align because of the initial positioning error without the CT-guided in Fig. 2(d1). After the conventional alignment in Fig. 2(d2), the errors of the anatomy mapping in the CVs are reduced, but still do not match well due to the metal image artifact. In Fig. 2(d3), the overlay regions demonstrated that the anatomy aligns really well with the proposed method.

Figure 2.

Examples of CV mapping results on the CBCT image of an HNC patient. Rows (1), (2) and (3) show results without alignment, with conventional global alignment, and with the proposed method, respectively. Columns (a) and (b) show checkerboard displays of two axial slices. Column (c) shows superposition of 3D renderings. Column (d) shows the selected yellow rectangular CVs.

Figure 3 shows the 5% and 95% Box-and-whisker plots for the differences in couch shifts between the measurement and reference for the testing dataset in the six dimensions. It is evident that the proposed method achieves accurate positioning for the HNC patient, and the proposed method outperforms the traditional method. In Table 1, the following observations can be made: first, the system positioning errors of translation and rotation are less than 0.47 mm and 0.17 degree, respectively. The random positioning errors of translation and rotation are less than 1.13 mm and 0.29 degree, respectively. Second, the percentage of testing fractions that fell below a predetermined tolerance (±2.0mm/±1.0°) with the proposed method is 87.88% and 100.00% in translations and rotations respectively. The accepted percentage of HNC cases has improved from 66.67% for the conventional method to 90.91 % for the proposed method.

Figure 3.

The 5% and 95% Box-and-whisker plots for the differences in couch shifts between the measurement and reference. The blue and red boxes indicate the measurement using the conventional and the proposed method, respectively. Abbreviation: A-P = anterior-posterior, L-R = left-right, S-I = superior-inferior.

Table 1.

Summary of the errors with different methods. We measure the proportion of the fractions within a predetermined tolerance (±2mm/±1.0º) for each catalog, overall mean M, random error σ, and systematic error ∑. Note that the execution time does not include the time spent on annotation of control volumes.

Metric	Method	Translation (mm)			Rotation (°)
		A-P	L-R	S-I	Yaw	Pitch	Roll
Overall mean (M)	Conventional	0.33	−0.25	0.45	−0.36	0.11	0.12
	DL without CVs	−0.36	−0.66	0.36	−0.26	0.07	0.15
	DL with CVs	−0.19	0.43	0.20	−0.05	−0.03	0.10
Systematic error (∑)	Conventional	0.71	1.08	1.03	0.16	0.26	0.30
	DL without CVs	0.67	1.40	0.66	0.35	0.40	0.25
	DL with CVs	0.44	0.27	0.47	0.11	0.17	0.10
Random error (σ)	Conventional	1.47	2.02	1.78	0.43	0.56	0.69
	DL without CVs	1.42	1.69	1.25	0.76	0.89	0.55
	DL with CVs	1.13	0.61	0.88	0.23	0.29	0.23
Within 2.0mm/1.0° (%)	Conventional	78.79	72.73	84.85	90.91	96.97	87.88
	DL without CVs	75.76	75.76	87.88	90.91	93.94	87.88
	DL with CVs	87.88	100.00	96.97	100.00	100.00	100.00
Within 2.0mm/1.0° in all direction (%)	Conventional	66.67
	DL without CVs	63.64
	DL with CVs	90.91
Execution Time for per fraction (ms)	Conventional	1919 ± 187
	DL without CVs	201 ± 69
	DL with CVs	214 ± 72

Moreover, we have also added quantitative comparison of the deep learning-based registration results obtained with and without control volumes in Table 1. With the incorporation of the CVs, we found that 90.91 % of the cases are clinically acceptable, which significantly improves the acceptable proportion of 63.6 % for DL-based registration without inclusion of the CVs. To evaluate the mapping accuracies of the CVs, we determined the landmarks mapping errors separately for individual CV. The mapping accuracy of individual CV in the testing dataset is shown in figure 4 using the proposed method. The mean error of the landmarks mapping is less than 2.5 mm in all the CVs.

Figure 4.

The 5% and 95% Box-and-whisker plots for the landmarks mapping error between the proposed method and the reference of the 4 CVs.

The value for λ was determined empirically by analyzing the global image and the CVs’ NCC value while keeping other parameters fixed for the validation set. The most ideal selection scheme for λ is that the NCC is maximum in both global image and CV. However, because of the deformation of the head and neck, the case where NCC is maximum in global image and CV is not the same λ. Figure 5 shows that NCC has a better balance between global image and CV when λ = 10. Therefore, we set the λ parameter to 10.

Figure 5.

NCC of the global image and the CVs on the validation dataset with different λ during the network validation.

4. DISCUSSION

Automatic patient positioning is one of the most challenging tasks and quality issues in HNC patient RT. Due to limitations of traditional registration algorithms, HNC patient positioning commonly requires manual registration. Different from the previous patient positioning strategies, in this study, a novel unsupervised DL strategy of automated CVs mapping on dCT is investigated. The novel aspects of our method are as follows: First, an unsupervised DL architecture with inclusion of CVs mapping is proposed, which weights the CVs regions differently to mitigate any potential adverse influence of image artifacts on the registration. Second, we introduce a loss function that considers the CV-local similarity with respect to the transformational parameters. The trained network enables to mitigate the effect of image artifacts in order to achieve automatic human-level comparable HNC patient positioning.

In theory, positioning shifts may be difference in position between tumor target at planning and treatment [17]. In reality, most of the tumor targets are soft tissues and may be difficult to visualize. In addition, tumors may change during RT [6, 41]. Thus, the adjacent bony anatomy is often used as surrogate for alignment in HNC RT [42]. It’s possible that the choice of reference bony structures for calculating setup deviations is arbitrary and varies in reality. The introduction of CVs offers a convenient technique that can map the arbitrary bony structure between the dCT and pCT. The unsupervised DL model can also be applied to map the other anatomies encompassed by CVs to meet different clinical demands.

The individual bony anatomy is a rigid object. However, the complex anatomy of the whole HNC area must be considered as a non-rigid object. Implementing a rigid alignment for the entire bony anatomy may return an ambiguous result because even after a successful overall alignment, not all structures will be well aligned [41]. This situation is difficult to distinguish from a failed alignment. To address this, we introduced a technique that mapped separately multiple CVs and generated the transformational parameters for each individual CV. The final single couch shift was derived as the average of all transformational parameters based on the individual CVs.

In the training stage, the CV is randomly selected to achieve unsupervised learning without any manual annotation. The network tries to map the arbitrary CV of the head and neck from daily CT to planning CT in the training stage. The proposed method offers a tool for physicians to map all of the interests of volume. We would like to clarify that the placement (or “annotation”) of the CVs in the testing stage is only manually done in the planning CT, not the daily CT. The annotation/segmentation of the multi-target is a routine process during treatment planning. Thus the “annotation” of a CV is rather straightforward, involving simply the placement of the CV in a bony anatomic region in a fairly casual fashion (this CV placement process can be easily automated). Clinically, no human intervention is involved during daily CT acquisition and patient positioning, thus it does not slow patient repositioning.

Recent research shows that image registration using a semi-supervised learning method assisted by the manual segmentation can effectively improve the accuracy of registration [43, 44]. In future research, we plan to manually segment the target region of interest in the training data. The image registration based on semi-supervised learning is realized by using the loss function which combines the image similarity and segmentation accuracy measurement.

Further improvements may be carried out in the future as follows. First, we adopted the 3D network in unsupervised DL model, which was time-consuming. The training step’s efficiency can be enhanced by balancing network width, depth, and resolution [45]. Second, insufficient training images may affect the accuracy of the results. To enhance the robustness of the network, we will further enlarge the datasets with a large variety of anatomic variations and from different CT scanners in the future. In this paper, we enrolled 158 patients with 554 CT scans for model evaluation. Enlarging the training dataset is expected to further improve the accuracy of the HNC patient positioning through avoiding potential overfitting. In order to obtain more convincing results, the cross-validation will be conducted in the future research. Third, multiple factors can contribute to the heterogeneity in errors, such as tumor shrinkage, severity of artifacts, learned versus test performance, etc.

The DL models of medical imaging are poorly interpretable, and their “black-box” properties make it difficult for users to fully understand the decision basis of the models. It is also difficult for existing deep learning models to provide model corrections based on prior knowledge. The interpretability and visualization of DL models will be key for artificial intelligence to gain the trust of physicians and regulators in the future.

5. CONCLUSION

In conclusion, we proposed an unsupervised DL framework for highly accurate and efficient patient positioning. In addition to image correlation between pCT and dCT, we incorporated the CV mapping to improve the robustness of the registration and alleviate the effect of the image artifact. To our knowledge, this represents the new attempt to use unsupervised DL to HNC CT-guided RT. The acceptable percentage of cases increased from 66.67% for the conventional method to 90.91 % for the proposed method. The concept of our proposed method, coupled with promising initial demonstration, suggests that the proposed method has a strong potential for becoming a valuable IGRT tool.

• An unsupervised deep learning framework is proposed for highly accurate and efficient patient positioning in CT-guided radiotherapy.

• We incorporate a CVs mapping strategy into the DL architecture, which emphasizes the CVs regions to mitigate any potential adverse influence of image artifacts on registration.