Deep learning based automatic segmentation of the Internal Pudendal Artery in definitive radiotherapy treatment planning of localized prostate cancer.

Background and purpose: Radiation-induced erectile dysfunction (RiED) commonly affects prostate cancer patients, prompting clinical trials across institutions to explore dose-sparing to internal-pudendal-arteries (IPA) for preserving sexual potency. IPA, challenging to segment, isn't conventionally considered an organ-at-risk (OAR). This study proposes a deep learning (DL) auto-segmentation model for IPA, using Computed Tomography (CT) and Magnetic Resonance Imaging (MRI) or CT alone to accommodate varied clinical practices. Materials and methods: A total of 86 patients with CT and MRI images and noisy IPA labels were recruited in this study. We split the data into 42/14/30 for model training, testing, and a clinical observer study, respectively. There were three major innovations in this model: 1) we designed an architecture with squeeze-and-excite blocks and modality attention for effective feature extraction and production of accurate segmentation, 2) a novel loss function was used for training the model effectively with noisy labels, and 3) modality dropout strategy was used for making the model capable of segmentation in the absence of MRI. Results: Test dataset metrics were DSC 61.71 ± 7.7 %, ASD 2.5 ± .87 mm, and HD95 7.0 ± 2.3 mm. AI segmented contours showed dosimetric similarity to expert physician's contours. Observer study indicated higher scores for AI contours (mean = 3.7) compared to inexperienced physicians' contours (mean = 3.1). Inexperienced physicians improved scores to 3.7 when starting with AI contours. Conclusion: The proposed model achieved good quality IPA contours to improve uniformity of segmentation and to facilitate introduction of standardized IPA segmentation into clinical trials and practice.

A Proof-of-Concept Study of Artificial Intelligence-assisted Contour Editing.

Purpose: To present a concept called artificial intelligence-assisted contour editing (AIACE) and demonstrate its feasibility. Materials and Methods: The conceptual workflow of AIACE is as follows: Given an initial contour that requires clinician editing, the clinician indicates where large editing is needed, and a trained deep learning model uses this input to update the contour. This process repeats until a clinically acceptable contour is achieved. In this retrospective, proof-of-concept study, the authors demonstrated the concept on two-dimensional (2D) axial CT images from three head-and-neck cancer datasets by simulating the interaction with the AIACE model to mimic the clinical environment. The input at each iteration was one mouse click on the desired location of the contour segment. Model performance is quantified with the Dice similarity coefficient (DSC) and 95th percentile of Hausdorff distance (HD95) based on three datasets with sample sizes of 10, 28, and 20 patients. Results: The average DSCs and HD95 values of the automatically generated initial contours were 0.82 and 4.3 mm, 0.73 and 5.6 mm, and 0.67 and 11.4 mm for the three datasets, which were improved to 0.91 and 2.1 mm, 0.86 and 2.5 mm, and 0.86 and 3.3 mm, respectively, with three mouse clicks. Each deep learning-based contour update required about 20 msec. Conclusion: The authors proposed the newly developed AIACE concept, which uses deep learning models to assist clinicians in editing contours efficiently and effectively, and demonstrated its feasibility by using 2D axial CT images from three head-and-neck cancer datasets.Keywords: Segmentation, Convolutional Neural Network (CNN), CT, Deep Learning Algorithms Supplemental material is available for this article. © RSNA, 2022.

Site-agnostic 3D dose distribution prediction with deep learning neural networks.

PURPOSE: Typically, the current dose prediction models are limited to small amounts of data and require retraining for a specific site, often leading to suboptimal performance. We propose a site-agnostic, three-dimensional dose distribution prediction model using deep learning that can leverage data from any treatment site, thus increasing the total data available to train the model. Applying our proposed model to a new target treatment site requires only a brief fine-tuning of the model to the new data and involves no modifications to the model input channels or its parameters. Thus, it can be efficiently adapted to a different treatment site, even with a small training dataset. METHODS: This study uses two separate datasets/treatment sites: data from patients with prostate cancer treated with intensity-modulated radiation therapy (source data), and data from patients with head-and-neck cancer treated with volumetric-modulated arc therapy (target data). We first developed a source model with 3D UNet architecture, trained from random initial weights on the source data. We evaluated the performance of this model on the source data. We then studied the generalizability of the model to the new target dataset via transfer learning. To do this, we built three more models, all with the same 3D UNet architecture: target model, adapted model, and combined model. The source and target models were trained on the source and target data from random initial weights, respectively. The adapted model fine-tuned the source model to the target domain by using the target data. Finally, the combined model was trained from random initial weights on a combined data pool consisting of both target and source datasets. We tested all four models on the target dataset and evaluated quantitative dose-volume histogram metrics for the planning target volume (PTV) and organs at risk (OARs). RESULTS: When tested on the source treatment site, the source model accurately predicted the dose distributions with average (mean, max) absolute dose errors of (0.32%±0.14, 2.37%±0.93) (PTV) relative to the prescription dose, and highest mean dose error of 1.68%±0.76, and highest max dose error of 5.47%± 3.31 for femoral head right. The error in PTV dose coverage prediction is 3.21%±1.51 for D98 , 3.04%±1.69 for D95 , and 1.83%±1.01 for D02 . Averaging across all OARs, the source model predicted the OAR mean dose within 1.38% and the OAR max dose within 3.64%. For the target treatment site, the target model average (mean, max) absolute dose errors relative to the prescription dose for the PTV were (1.08%±0.95, 2.90%±1.35). Left cochlea had the highest mean and max dose errors of 5.37%±5.82 and 8.33%±8.88, respectively. The errors in PTV dose coverage prediction for D98 and D95 were 2.88%±1.59 and 2.55%±1.28, respectively. The target model can predict the OAR mean dose within 2.43% and the OAR max dose within 4.33% on average across all OARs. CONCLUSION: We developed a site-agnostic model for three-dimensional dose prediction and tested its adaptability to a new target treatment site via transfer learning. Our proposed model can make accurate predictions with limited training data.

PSA-Net: Deep learning-based physician style-aware segmentation network for postoperative prostate cancer clinical target volumes.

PURPOSE: Automatic segmentation of medical images with deep learning (DL) algorithms has proven highly successful in recent times. With most of these automation networks, inter-observer variation is an acknowledged problem that leads to suboptimal results. This problem is even more significant in segmenting postoperative clinical target volumes (CTV) because they lack a macroscopic visible tumor in the image. This study, using postoperative prostate CTV segmentation as the test case, tries to determine 1) whether physician styles are consistent and learnable, 2) whether physician style affects treatment outcome and toxicity, and 3) how to explicitly deal with different physician styles in DL-assisted CTV segmentation to facilitate its clinical acceptance. METHODS: A dataset of 373 postoperative prostate cancer patients from UT Southwestern Medical Center was used for this study. We used another 83 patients from Mayo Clinic to validate the developed model and its adaptability. To determine whether physician styles are consistent and learnable, we trained a 3D convolutional neural network classifier to identify which physician had contoured a CTV from just the contour and the corresponding CT scan. Next, we evaluated whether adapting automatic segmentation to specific physician styles would be clinically feasible based on a lack of difference between outcomes. Here, biochemical progression-free survival (BCFS) and grade 3+ genitourinary and gastrointestinal toxicity were estimated with the Kaplan-Meier method and compared between physician styles with the log rank test and subsequently with a multivariate Cox regression. When we found no statistically significant differences in outcome or toxicity between contouring styles, we proposed a concept called physician style-aware (PSA) segmentation by developing an encoder-multidecoder network with perceptual loss to model different physician styles of CTV segmentation. RESULTS: The classification network captured the different physician styles with 87% accuracy. Subsequent outcome analysis showed no differences in BCFS and grade 3+ toxicity among physicians. With the proposed physician style-aware network (PSA-Net), Dice similarity coefficient (DSC) accuracy for all physicians was 3.4% higher on average than with a general model that does not differentiate physician styles. We show that these stylistic contouring variations also exist between institutions that follow the same segmentation guidelines, and we show the proposed method's effectiveness in adapting to new institutional styles. We observed an accuracy improvement of 5% in terms of DSC when adapting to the style of a separate institution. CONCLUSION: The performance of the classification network established that physician styles are learnable, and the lack of difference between outcomes among physicians shows that the network can feasibly adapt to different styles in the clinic. Therefore, we developed a novel PSA-Net model that can produce contours specific to the treating physician, thus improving segmentation accuracy and avoiding the need to train multiple models to achieve different style segmentations. We successfully validated this model on data from a separate institution, thus supporting the model's generalizability to diverse datasets.

A deep learning-based framework for segmenting invisible clinical target volumes with estimated uncertainties for post-operative prostate cancer radiotherapy.

In post-operative radiotherapy for prostate cancer, precisely contouring the clinical target volume (CTV) to be irradiated is challenging, because the cancerous prostate gland has been surgically removed, so the CTV encompasses the microscopic spread of tumor cells, which cannot be visualized in clinical images like computed tomography or magnetic resonance imaging. In current clinical practice, physicians' segment CTVs manually based on their relationship with nearby organs and other clinical information, but this allows large inter-physician variability. Automating post-operative prostate CTV segmentation with traditional image segmentation methods has yielded suboptimal results. We propose using deep learning to accurately segment post-operative prostate CTVs. The model proposed is trained using labels that were clinically approved and used for patient treatment. To segment the CTV, we segment nearby organs first, then use their relationship with the CTV to assist CTV segmentation. To ease the encoding of distance-based features, which are important for learning both the CTV contours' overlap with the surrounding OARs and the distance from their borders, we add distance prediction as an auxiliary task to the CTV network. To make the DL model practical for clinical use, we use Monte Carlo dropout (MCDO) to estimate model uncertainty. Using MCDO, we estimate and visualize the 95% upper and lower confidence bounds for each prediction which informs the physicians of areas that might require correction. The model proposed achieves an average Dice similarity coefficient (DSC) of 0.87 on a holdout test dataset, much better than established methods, such as atlas-based methods (DSC<0.7). The predicted contours agree with physician contours better than medical resident contours do. A reader study showed that the clinical acceptability of the automatically segmented CTV contours is equal to that of approved clinical contours manually drawn by physicians. Our deep learning model can accurately segment CTVs with the help of surrounding organ masks. Because the DL framework can outperform residents, it can be implemented practically in a clinical workflow to generate initial CTV contours or to guide residents in generating these contours for physicians to review and revise. Providing physicians with the 95% confidence bounds could streamline the review process for an efficient clinical workflow as this would enable physicians to concentrate their inspecting and editing efforts on the large uncertain areas.

A comparison of Monte Carlo dropout and bootstrap aggregation on the performance and uncertainty estimation in radiation therapy dose prediction with deep learning neural networks.

Recently, artificial intelligence technologies and algorithms have become a major focus for advancements in treatment planning for radiation therapy. As these are starting to become incorporated into the clinical workflow, a major concern from clinicians is not whether the model is accurate, but whether the model can express to a human operator when it does not know if its answer is correct. We propose to use Monte Carlo Dropout (MCDO) and the bootstrap aggregation (bagging) technique on deep learning (DL) models to produce uncertainty estimations for radiation therapy dose prediction. We show that both models are capable of generating a reasonable uncertainty map, and, with our proposed scaling technique, creating interpretable uncertainties and bounds on the prediction and any relevant metrics. Performance-wise, bagging provides statistically significant reduced loss value and errors in most of the metrics investigated in this study. The addition of bagging was able to further reduce errors by another 0.34% for [Formula: see text] and 0.19% for [Formula: see text] on average, when compared to the baseline model. Overall, the bagging framework provided significantly lower mean absolute error (MAE) of 2.62, as opposed to the baseline model's MAE of 2.87. The usefulness of bagging, from solely a performance standpoint, does highly depend on the problem and the acceptable predictive error, and its high upfront computational cost during training should be factored in to deciding whether it is advantageous to use it. In terms of deployment with uncertainty estimations turned on, both methods offer the same performance time of about 12 s. As an ensemble-based metaheuristic, bagging can be used with existing machine learning architectures to improve stability and performance, and MCDO can be applied to any DL models that have dropout as part of their architecture.

Fully automated organ segmentation in male pelvic CT images.

Accurate segmentation of prostate and surrounding organs at risk is important for prostate cancer radiotherapy treatment planning. We present a fully automated workflow for male pelvic CT image segmentation using deep learning. The architecture consists of a 2D organ volume localization network followed by a 3D segmentation network for volumetric segmentation of prostate, bladder, rectum, and femoral heads. We used a multi-channel 2D U-Net followed by a 3D U-Net with encoding arm modified with aggregated residual networks, known as ResNeXt. The models were trained and tested on a pelvic CT image dataset comprising 136 patients. Test results show that 3D U-Net based segmentation achieves mean (±SD) Dice coefficient values of 90 (±2.0)%, 96 (±3.0)%, 95 (±1.3)%, 95 (±1.5)%, and 84 (±3.7)% for prostate, left femoral head, right femoral head, bladder, and rectum, respectively, using the proposed fully automated segmentation method.

Hepatocellular Carcinoma Tumor Dose Response After 90Y-radioembolization With Glass Microspheres Using 90Y-SPECT/CT-Based Voxel Dosimetry.

PURPOSE: To investigate hepatocellular carcinoma tumor dose-response characteristics based on voxel-level absorbed doses (D) and biological effective doses (BED) using quantitative 90Y-single-photon emission computed tomography (SPECT)/computed tomography (CT) after 90Y-radioembilization with glass microspheres. We also investigated the relationship between normal liver D and toxicities. METHODS AND MATERIALS: 90Y-radioembolization activity distributions for 34 patients were based on quantitative 90Y-bremsstrahlung SPECT/CT. D maps were generated using a local-deposition algorithm. Contrast-enhanced CT or magnetic resonance imaging scans of the liver were registered to 90Y-SPECT/CT, and all tumors larger than 2.5 cm diameter (53 tumors) were segmented. Tumor mean D and BED (Dmean and BEDmean) and dose volume coverage from 0% to 100% in 10% steps (D0-D100 and BED0-BED100) were extracted. Tumor response was evaluated on follow-up using World Health Organization (WHO), Response Evaluation Criteria in Solid Tumors (RECIST), and modified RECIST (mRECIST) criteria. Differences in dose metrics for responders and nonresponders were assessed using the Mann-Whitney U test. A univariate logistic regression model was used to determine tumor dose metrics that correlated with tumor response. Correlations among tumor size, tumor Dmean, and tumor dose heterogeneity (defined as the coefficient of variation) were assessed. RESULTS: The objective response rates were 14 of 53, 15 of 53, and 30 of 53 for WHO, RECIST, and mRECIST criteria, respectively. WHO and RECIST response statuses did not correlate with D or BED. For mRECIST responders and nonresponders, D and BED were significantly different for Dmean, D20 to D80, BEDmean, and BED0 to BED80. Threshold doses (and the 95% confidence interval) for 50% probability of mRECIST response (D50%) were 160 Gy (123-196 Gy) for Dmean and 214 Gy (146-280 Gy) for BEDmean. Tumor dose heterogeneity significantly correlated with tumor volume. No statistically significant association between Dmean to normal liver and complications related to bilirubin, albumin, or ascites was observed. CONCLUSIONS: Hepatocellular carcinoma tumor dose-response curves after 90Y-radioembolization with glass microspheres showed Dmean of 160 Gy and BEDmean of 214 Gy for D50% with a positive predictive value of ∼70% and a negative predictive value of ∼62%. No complications were observed in our patient cohort for normal liver Dmean less than 44 Gy.

Characterization of 90 Y-SPECT/CT self-calibration approaches on the quantification of voxel-level absorbed doses following 90 Y-microsphere selective internal radiation therapy.

PURPOSE: 90 Y-microsphere selective internal radiation therapy (90 Y-SIRT or 90 Y-radioembolization) is used in the management of unresectable liver tumors. 90 Y-SIRT presents a unique situation where the total 90 Y activity inside the liver can be determined with high accuracy (> 95%). 90 Y bremsstrahlung single-photon emission computed tomography (SPECT)/computed tomography (CT) can be self-calibrated to provide quantitative images that facilitate voxel-level absorbed dose calculations. We investigated the effects of different approaches for 90 Y-SPECT self-calibration on the quantification of absorbed doses following 90 Y-SIRT. METHODS: 90 Y bremsstrahlung SPECT/CT images of 31 patients with hepatocellular carcinoma, collected following 90 Y-SIRT, were analyzed, yielding 48 tumor and 31 normal liver contours. We validated the accuracy of absorbed doses calculated by a commercial software against those calculated using Monte Carlo-based radiation transport. The software package was used to analyze the following definitions of SPECT volume of interest used for 90 Y-SPECT self-calibration: (a) SPECT field-of-view (FOV), (b) chest-abdomen contour, (c) total liver contour, (d) total liver contour expanded by 5 mm, and (e) total liver contour contracted by 5 mm. Linear correlation and Bland-Altman analysis were performed for tumor and normal liver tissue absorbed dose volume histogram metrics between the five different approaches for 90 Y-SPECT self-calibration. RESULTS: The mean dose calculated using the commercial software was within 3% of Monte Carlo for tumors and normal liver tissues. The tumor mean dose calculated using the chest-abdomen calibration was within 2% of that calculated using the SPECT FOV, whereas the doses calculated using the total liver contour, expanded total liver contour, and contracted total liver contour were within 68%, 47%, and 107%, respectively, of doses calculated using the SPECT FOV. The normal liver tissue mean dose calculated using the chest-abdomen contour was within 1.3% of that calculated using the SPECT FOV, whereas the doses calculated using the total liver contour, expanded total liver contour, and contracted total liver contour were within 73%, 50%, and 114%, respectively, of doses calculated using the SPECT FOV. CONCLUSIONS: The mean error of < 3% for commercial software can be considered clinically acceptable for 90 Y-SIRT dosimetry. Absorbed dose quantification using 90 Y-SPECT self-calibration with the chest-abdomen contour was equivalent to that calculated using the SPECT FOV, but self-calibration with the total liver contour yielded substantially higher (~70%) dose values. The large biases revealed by our study suggest that consistent absorbed dose calculation approaches are essential when comparing 90 Y-SIRT dosimetry between different clinical studies.