Radiation dose to specific cardiac substructures can have a significant on treatment related morbidity and mortality, yet definition of these structures is labor intensive and not standard. Autosegmentation software may potentially address these issues, however it is unclear whether this approach can be broadly applied across different treatment planning conditions. We investigated the feasibility of autosegmentation of the cardiac substructures in four-dimensional (4D) computed tomography (CT), respiratory-gated, non-contrasted imaging.

To determine whether autosegmentation can be successfully employed on 4DCT respiratory-gated, non-contrasted imaging.

We included patients who underwent stereotactic body radiation therapy for inoperable, early-stage non-small cell lung cancer from 2007 to 2019. All patients were simulated via 4DCT imaging with respiratory gating without intravenous contrast. Generated structure quality was evaluated by degree of required manual edits and volume discrepancy between the autocontoured structures and its edited sister structure.

Initial 17-structure cardiac atlas was generated with 20 patients followed by three successive iterations of 10 patients using MIM software. The great vessels and heart chambers were reliably autosegmented with most edits considered minor. In contrast, coronary arteries either failed to be autosegmented or the generated structures required major alterations necessitating deletion and manual definition. Similarly, the generated mitral and tricuspid valves were poor whereas the aortic and pulmonary valves required at least minor and moderate changes respectively. For the majority of subsites, the additional samples did not appear to substantially impact the quality of generated structures. Volumetric analysis between autosegmented and its manually edited sister structure yielded comparable findings to the physician-based assessment of structure quality.

The use of MIM software with 30-sample subject library was found to be useful in delineating many of the heart substructures with acceptable clinical accuracy on respiratory-gated 4DCT imaging. Small volume structures, such as the coronary arteries were poorly autosegmented and require manual definition.

Cardiotoxicity is a significant concern in modern radiotherapy (RT). RTOG 0617 was a large, phase III trial investigating the utility of dose escalation in locally advanced non-small cell lung cancer (NSCLC), where the experimental dose escalated arm reduced patient survival, in part due to excessive cardiotoxicity[1]. In a follow up study, Thor et al[2] revealed that dose to specific cardiac substructures such as the atria, ventricles, and pericardium were highly correlated with overall survival. These findings have been supported by several other groups, who also demonstrate a dose-dependent relationship between cardiac substructures and outcome[3,4]. Despite these findings, definition of such structures is not standard in radiation planning at this time.

Conventionally, the heart is contoured as a single structure which encompasses the atria, ventricles, valve, pericardium, and coronary arteries. Individual definition and subsequent dose calculation to the cardiac substructures is not routinely performed. Delineation of cardiac subsites can be challenging due to heart and respiratory motion, lack of intravenous contrast on treatment planning computed tomography (CT) imaging, and anatomic variances[5-10]. Furthermore, contouring these structures can be time consuming and not necessarily practical in typical care[10].

Researchers have attempted to overcome these obstacles utilizing a number of approaches. For example, several groups have provided cardiac atlases for these subsites, including the recently closed clinical trial RTOG 1106 (NCT01507428) to provide guidance on how to define these structures[5,10,11]. Additionally, there is evidence that autosegmentation is a feasible option to mitigate the added labor associated with including these structures[6-10,12]. While this is a provocative strategy, this method may not be viable in all treatment conditions. For example, four-dimensional (4D) CT imaging with respiratory gating is a treatment delivery technique for stereotactic body radiation therapy (SBRT) utilized by approximately 31%-54% of providers in the United States[13,14]. As this requires multiple scans to construct each phase of respiration, timed intravenous contrast is not possible and cardiac motion may be amplified. Therefore, it is unclear how well autosegmentation would perform in this setting. We investigated whether autosegmentation for cardiac substructures could be accurately employed on non-contrasted, respiratory gated imaging.

Utilizing MIM software, cardiac substructures were autosegmented in 10 patients and these structures were then manually refined (Figure 1). This process then repeated two subsequent times and representative autosegmented structures following each iteration can be found in Supplementary Figures 2-4. The great vessels including the whole heart were reliably autosegmented with the majority requiring minor changes without a substantial improvement from additional samples (Figure 2). In contrast, coronary arteries either failed to be autosegmented or the generated structures required major alterations necessitating deletion and manual definition. Similarly, the generated mitral and tricuspid valves were poor whereas the aortic and pulmonary valves required at least minor and moderate changes respectively (Figure 2). The heart chambers, particularly the atria, were autocontoured well (Figure 2). As compared to structures generated with 20 samples, there was a general improvement in structure quality after 30 samples, however, this trend did not continue in the 40-sample group.

Open in New Tab   Full Size Figure   Download Figure

Figure 2 Physician-based evaluation of generated structures using atlas libraries comprised of 20-, 30-, and 40-patient samples (mean, standard deviation). Scale: 0 = No changes made; 1 = Minor changes made, 2 = Moderate changes required by structure was salvageable, 3 = Structure required deletion and replacement, 4 = Structure failed autosegmentation.

Additionally, volumes of the manually contoured and autosegmented cardiac substructures for the respective groups were compared. In general, MIM undercontoured structures when compared to the manually derived contours (Figure 3). This discrepancy was typically smaller for cardiac substructures with larger volumes such as the great vessels and heart chambers, whereas the difference was more profound in the smaller structures (Figure 3). There was excellent agreement in the volume of autogenerated and manual structures for the whole heart (Supplementary Figure 5). In some structures (right ventricle, descending aorta, pulmonary artery), more samples improved the volume discrepancy between manual and autocontoured structures (Figure 3). Unfortunately, often the autosegmented structures for the coronary arteries, mitral, and tricuspid valves were very small (< 0.1 cm³) and essentially unusable.

Open in New Tab   Full Size Figure   Download Figure

Figure 3 Analysis comparing the volumes of autosegmented structures (MIM) vs subsequent physician-edited sister (RO) (mean ± SD).

In current study, we explored the feasibility of MIM autocontouring for cardiac substructures on respiratory gated, non-contrasted 4DCT scans. Following autosegmentation, the great vessels and heart chambers required minor to no changes on average, whereas coronary arteries, tricuspid and mitral valves often failed autosegmentation or required complete recontouring. The aortic and pulmonary valves required a greater degree of modification but ultimately were usable structures. Similar observations were seen on volumetric analysis between the autosegmented structures and their manually edited correlates.

Several atlases for cardiac substructures have been described. The current study defined heart subsites per Feng et al[5], who used contrast-enhanced CT images of breast cancer patients to provide a well-defined atlas which is similar to previous studies, including RTOG 1106 (NCT01507428)[5,11,16].

Prior reports have evaluated methods for autocontouring cardiac substructures. Kirisli et al[17] utilized a multi-atlas approach to autosegment the whole heart and cardiac chambers using CT angiography. Shahzad et al[9] relied on contrast enhanced imaging for atlas generation, yet was then able to successfully autosegment structures on non-contrasted imaging via this template.

Others have demonstrated the reliability of non-contrasted images for atlas construction. In a breast cancer population, Finnegan et al[6] used multi-atlas segmentation in non-contrasted, free-breath CT imaging reporting similar success with the great vessels, whole heart, and heart chambers. Moreover, Luo et al[8] utilized non-contrasted, 4DCT images for multi-atlas autosegmentation finding minimal changes needed in generated structures with no dosimetric difference between autosegmented and manually derived structures. Additional groups have successfully used 4DCT imaging in deriving cardiac atlases and subsequent autosegmentation[3,10]. Consistent with our findings, these studies have reported success in autocontouring the great vessels and heart chambers, whereas the coronary arteries and heart valves have too much variability to be reliably used, despite a wide variety of planning techniques[3,6,8,10].

In the current report, the quality and volume agreement of the studied structures improved when using a 30 vs 20 patient atlas, however this trend did not continue when 40 patients were used. The reason for this discrepancy is likely due to anatomic variances. The study cohort consisted of patients planned for intrathoracic SBRT, a group that often has a high proportion of cardiac disease (manuscript in press). As such, it is not uncommon to see cardiac hypertrophy, atherosclerosis, implantable hardware (pacemakers, cardiac stents), and other cardiovascular pathologies within this patient population. These factors can ultimately impact the ability to accurately autosegment heart substructures either due to the relative high hounsfield units of implants/calcium or differences compared to normal anatomy. Indeed, four patients within the 40-sample group who required a high degree of modification of autocontoured structures had such abnormalities (pacemaker and cardiac stents, 1 patient; sternotomy wires and atrial/ventricular hypertrophy, 1 patient; significant atherosclerosis and atrial/ventricular hypertrophy, 2 patients). Therefore, while increasing the sample number in the atlas database can improve the quality of generated structures, individual patient factors can still significantly impact autosegmentation.

There are several limitations in our study. The lack of intravenous contrast can limit the identification of cardiac substructures, particularly the coronary arteries and heart valves. However, given that these are small volume structures significantly impacted by cardiac motion and the difficulty that previous studies had with autosegmenting these subsites, contrast alone may not be sufficient to overcome this issue. Additionally, the conducting pathways of the heart cannot be reliably identified on CT[5]. Future studies incorporating cardiac magnetic resonance imaging could be useful in this setting. Lastly, while increasing the number of patients in the MIM atlas did not consistently improve contour quality in this study, given the limited number patients it is unclear whether this is applicable to much larger datasets. Strengths of this study includes the use of respiratory gated planning CTs, as this technique is commonly used for motion management in intrathoracic SBRT. While respiratory gating can amplify cardiac motion, the current study demonstrates that MIM autosegmentation is viable for the great vessels and heart chambers thus enabling a practical approach to dose calculation to these structures. Future studies will use this approach to evaluate radiation dose to cardiac substructures during SBRT.

Accurate segmentation of various organs in the heart is necessary to understand and correlate cardiac toxicities in the management of NSCLC lung cancer patients being treated using SBRT dose delivery technique. Even though human heart is associated with non-voluntary motion, thereby, exhibiting motion artifacts during imaging, use of MIM software with 30-sample subject library was found to be useful in delineating majority of the heart substructures with acceptable clinical accuracy except some structures having very little volume. Moreover, the autosegmented structures need to be evaluated on case by case basis by the treating physician as presence of calcifications or other sources of imaging artifacts like pacemaker, etc., may inhibit the model to accurately auto segment the heart sub-structures.