Medical Physics

Radiation dose to cardiac substructures is related to radiation‐induced heart disease. However, substructures are not considered in radiation therapy planning (RTP) due to poor visualization on CT. Therefore, we developed a novel deep learning (DL) pipeline leveraging MRI’s soft tissue contrast coupled with CT for state‐of‐the‐art cardiac substructure segmentation requiring a single, non‐contrast CT input.

Thirty‐two left‐sided whole‐breast cancer patients underwent cardiac T2 MRI and CT‐simulation. A rigid cardiac‐confined MR/CT registration enabled ground truth delineations of 12 substructures (chambers, great vessels (GVs), coronary arteries (CAs), etc.). Paired MRI/CT data (25 patients) were placed into separate image channels to train a three‐dimensional (3D) neural network using the entire 3D image. Deep supervision and a Dice‐weighted multi‐class loss function were applied. Results were assessed pre/post augmentation and post‐processing (3D conditional random field (CRF)). Results for 11 test CTs (seven unique patients) were compared to ground truth and a multi‐atlas method (MA) via Dice similarity coefficient (DSC), mean distance to agreement (MDA), and Wilcoxon signed‐ranks tests. Three physicians evaluated clinical acceptance via consensus scoring (5‐point scale).

The model stabilized in ~19 h (200 epochs, training error <0.001). Augmentation and CRF increased DSC 5.0 ± 7.9% and 1.2 ± 2.5%, across substructures, respectively. DL provided accurate segmentations for chambers (DSC = 0.88 ± 0.03), GVs (DSC = 0.85 ± 0.03), and pulmonary veins (DSC = 0.77 ± 0.04). Combined DSC for CAs was 0.50 ± 0.14. MDA across substructures was <2.0 mm (GV MDA = 1.24 ± 0.31 mm). No substructures had statistical volume differences (P > 0.05) to ground truth. In four cases, DL yielded left main CA contours, whereas MA segmentation failed, and provided improved consensus scores in 44/60 comparisons to MA. DL provided clinically acceptable segmentations for all graded patients for 3/4 chambers. DL contour generation took ~14 s per patient.

These promising results suggest DL poses major efficiency and accuracy gains for cardiac substructure segmentation offering high potential for rapid implementation into RTP for improved cardiac sparing.

Manual delineation of head and neck (H&N) organ‐at‐risk (OAR) structures for radiation therapy planning is time consuming and highly variable. Therefore, we developed a dynamic multiatlas selection‐based approach for fast and reproducible segmentation.

Our approach dynamically selects and weights the appropriate number of atlases for weighted label fusion and generates segmentations and consensus maps indicating voxel‐wise agreement between different atlases. Atlases were selected for a target as those exceeding an alignment weight called dynamic atlas attention index. Alignment weights were computed at the image level and called global weighted voting (GWV) or at the structure level and called structure weighted voting (SWV) by using a normalized metric computed as the sum of squared distances of computed tomography (CT)‐radiodensity and modality‐independent neighborhood descriptors (extracting edge information). Performance comparisons were performed using 77 H&N CT images from an internal Memorial Sloan‐Kettering Cancer Center dataset (N = 45) and an external dataset (N = 32) using Dice similarity coefficient (DSC), Hausdorff distance (HD), 95th percentile of HD, median of maximum surface distance, and volume ratio error against expert delineation. Pairwise DSC accuracy comparisons of proposed (GWV, SWV) vs single best atlas (BA) or majority voting (MV) methods were performed using Wilcoxon rank‐sum tests.

Both SWV and GWV methods produced significantly better segmentation accuracy than BA (P < 0.001) and MV (P < 0.001) for all OARs within both datasets. SWV generated the most accurate segmentations with DSC of: 0.88 for oral cavity, 0.85 for mandible, 0.84 for cord, 0.76 for brainstem and parotids, 0.71 for larynx, and 0.60 for submandibular glands. SWV’s accuracy exceeded GWV's for submandibular glands (DSC = 0.60 vs 0.52, P = 0.019).

The contributed SWV and GWV methods generated more accurate automated segmentations than the other two multiatlas‐based segmentation techniques. The consensus maps could be combined with segmentations to visualize voxel‐wise consensus between atlases within OARs during manual review.

To quantify the predictive uncertainty in infrared (IR)‐marker‐based dynamic tumor tracking irradiation (IR Tracking) with Vero4DRT (MHI‐TM2000) for lung cancer using logfiles.

A total of 110 logfiles for 10 patients with lung cancer who underwent IR Tracking were analyzed. Before beam delivery, external IR markers and implanted gold markers were monitored for 40 s with the IR camera every 16.7 ms and with an orthogonal kV x‐ray imaging subsystem every 80 or 160 ms. A predictive model [four‐dimensional (4D) model] was then created to correlate the positions of the IR markers (P_IR) with the three‐dimensional (3D) positions of the tumor indicated by the implanted gold markers (P_detect). The sequence of these processes was defined as 4D modeling. During beam delivery, the 4D model predicted the future 3D target positions (P_predict) from the P_IR in real‐time, and the gimbaled x‐ray head then tracked the target continuously. In clinical practice, the authors updated the 4D model at least once during each treatment session to improve its predictive accuracy. This study evaluated the predictive errors in 4D modeling (E_4DM) and those resulting from the baseline drift of P_IR and P_detect during a treatment session (E_BD). E_4DM was defined as the difference between P_predict and P_detect in 4D modeling, and E_BD was defined as the mean difference between P_predict calculated from P_IR in updated 4D modeling using (a) a 4D model created from training data before the model update and (b) an updated 4D model created from new training data.

The meanE_4DM was 0.0 mm with the exception of one logfile. Standard deviations of E_4DM ranged from 0.1 to 1.0, 0.1 to 1.6, and 0.2 to 1.3 mm in the left‐right (LR), anterior–posterior (AP), and superior–inferior (SI) directions, respectively. The median elapsed time before updating the 4D model was 13 (range, 2–33) min, and the median frequency of 4D modeling was twice (range, 2–3 times) per treatment session. E_BD ranged from −1.0 to 1.0, −2.1 to 3.3, and −2.0 to 3.5 mm in the LR, AP, and SI directions, respectively. E_BD was highly correlated with BD_detect in the LR (R = −0.83) and AP directions (R = −0.88), but not in the SI direction (R = −0.40). Meanwhile, E_BD was highly correlated with BD_IR in the SI direction (R = −0.67), but not in the LR (R = 0.15) or AP (R = −0.11) direction. If the 4D model was not updated in the presence of intrafractional baseline drift, the predicted target position deviated from the detected target position systematically.

Application of IR Tracking substantially reduced the geometric error caused by respiratory motion; however, an intrafractional error due to baseline drift of >3 mm was occasionally observed. To compensate forE_BD, the authors recommend checking the target and IR marker positions constantly and updating the 4D model several times during a treatment session.

To determine how best to time respiratory surrogate‐based tumor motion model updates by comparing a novel technique based on external measurements alone to three direct measurement methods.

Concurrently measured tumor and respiratory surrogate positions from 166 treatment fractions for lung or pancreas lesions were analyzed. Partial‐least‐squares regression models of tumor position from marker motion were created from the first six measurements in each dataset. Successive tumor localizations were obtained at a rate of once per minute on average. Model updates were timed according to four methods:never, respiratory surrogate‐based (when metrics based on respiratory surrogate measurements exceeded confidence limits), error‐based (when localization error ≥3 mm), and always (approximately once per minute).

Radial tumor displacement prediction errors (mean ± standard deviation) for the four schema described above were 2.4 ± 1.2, 1.9 ± 0.9, 1.9 ± 0.8, and 1.7 ± 0.8 mm, respectively. Thenever‐update error was significantly larger than errors of the other methods. Mean update counts over 20 min were 0, 4, 9, and 24, respectively.

The same improvement in tumor localization accuracy could be achieved through any of the three update methods, but significantly fewer updates were required when therespiratory surrogate method was utilized. This study establishes the feasibility of timing image acquisitions for updating respiratory surrogate models without direct tumor localization.

Efficient compression of images while preserving image quality has the potential to be a major enabler of effective remote clinical diagnosis and treatment, since poor Internet connection conditions are often the primary constraint in such services. This paper presents a framework for organ‐specific image compression for teleinterventions based on a deep learning approach and anisotropic diffusion filter.

The proposed method, deep learning and anisotropic diffusion (DLAD), uses a convolutional neural network architecture to extract a probability map for the organ of interest; this probability map guides an anisotropic diffusion filter that smooths the image except at the location of the organ of interest. Subsequently, a compression method, such as BZ2 and HEVC‐visually lossless, is applied to compress the image. We demonstrate the proposed method on three‐dimensional (3D) CT images acquired for radio frequency ablation (RFA) of liver lesions. We quantitatively evaluate the proposed method on 151 CT images using peak‐signal‐to‐noise ratio (), structural similarity (), and compression ratio () metrics. Finally, we compare the assessments of two radiologists on the liver lesion detection and the liver lesion center annotation using 33 sets of the original images and the compressed images.

The results show that the method can significantly improve of most well‐known compression methods. DLAD combined with HEVC‐visually lossless achieves the highest average of 6.45, which is 36% higher than that of the original HEVC and outperforms other state‐of‐the‐art lossless medical image compression methods. The means of and are 70 dB and 0.95, respectively. In addition, the compression effects do not statistically significantly affect the assessments of the radiologists on the liver lesion detection and the lesion center annotation.

We thus conclude that the method has a high potential to be applied in teleintervention applications.

The importance of four‐dimensional‐magnetic resonance imaging (4D‐MRI) is increasing in guiding online plan adaptation in thoracic and abdominal radiotherapy. Many 4D‐MRI sequences are based on multislice two‐dimensional (2D) acquisitions which provide contrast flexibility. Intrinsic to MRI, however, are machine‐ and subject‐related geometric image distortions. Full correction of slice‐based 4D‐MRIs acquired on the Unity MR‐linac (Elekta AB, Stockholm, Sweden) is challenging, since through‐plane corrections are currently not available for 2D sequences. In this study, we implement a full three‐dimensional 3D correction and quantify the geometric and dosimetric effects of machine‐related (residual) geometric image distortions.

A commercial three‐dimensional (3D) geometric QA phantom (Philips, Best, the Netherlands) was used to quantify the effect of gradient nonlinearity (GNL) and static‐field inhomogeneity (B0I) on geometric accuracy. Additionally, the effectiveness of 2D (in‐plane, machine‐generic), 3D (machine‐generic), and in‐house developed 3D (machine‐specific) corrections was investigated. Corrections were based on deformable vector fields derived from spherical harmonics coefficients. Three patients with oligometastases in the liver were scanned with axial 4D‐MRIs on our MR‐linac (total: 10 imaging sessions). For each patient, a step‐and‐shoot IMRT plan (3 × 20 Gy) was created based on the simulation mid‐position (midP)‐CT. The 4D‐MRIs were then warped into a daily midP‐MRI and geometrically corrected. Next, the treatment plan was adapted according to the position offset of the tumor between midP‐CT and the 3D‐corrected midP‐MRIs. The midP‐CT was also deformably registered to the daily midP‐MRIs (different corrections applied) to quantify the dosimetric effects of (residual) geometric image distortions.

Using phantom data, median GNL distortions were 0.58 mm (no correction), 0.42–0.48 mm (2D), 0.34 mm (3D), and 0.34 mm (3D), measured over a diameter of spherical volume (DSV) of 200 mm. Median B0I distortions were 0.09 mm for the same DSV. For DSVs up to 500 mm, through‐plane corrections are necessary to keep the median residual GNL distortion below 1 mm. 3D and 3D corrections agreed within 0.15 mm. 2D‐corrected images featured uncorrected through‐plane distortions of up to 21.11 mm at a distance of 20–25 cm from the machine’s isocenter. Based on the 4D‐MRI patient scans, the average external body contour distortions were 3.1 mm (uncorrected) and 1.2 mm (2D‐corrected), with maximum local distortions of 9.5 mm in the uncorrected images. No (residual) distortions were visible for the metastases, which were all located within 10 cm of the machine’s isocenter. The interquartile range (IQR) of dose differences between planned and daily dose caused by variable patient setup, patient anatomy, and online plan adaptation was 1.37 Gy/Fx for the PTV D95%. When comparing dose on 3D‐corrected with uncorrected (2D‐corrected) images, the IQR was 0.61 (0.31) Gy/Fx.

GNL is the main machine‐related source of image distortions on the Unity MR‐linac. For slice‐based 4D‐MRI, a full 3D correction can be applied after respiratory sorting to maximize spatial fidelity. The machine‐specific 3D correction did not substantially reduce residual geometric distortions compared to the machine‐generic 3D correction for our MR‐linac. In our patients, dosimetric variations in the target not related to geometric distortions were larger than those caused by geometric distortions.