Development and validation of an AI-based lung lobe auto-contouring tool using radiation therapy planning free-breathing images Authors: Peter Ciaccio, Joseph Lombardo, Andrew Fuquay, Soroush Pahlavian, Rachel Grimm, Jun Kang, Wookjin Choi, Paul Sullivan, Yevgeniy Vinogradskiy Published: 2025 ABSTRACT BackgroundPulmonary toxicity can occur during radiation therapy of the lungs. Dose metrics evaluated at the lobar level can improve the ability to predict toxicity. Contouring lung lobes is challenging and time consuming. Currently there are limited dosimetry studies evaluating the dose to lung lobes. The purpose of this work was to develop and validate an artificial intelligence (AI) lung lobe auto-contouring algorithm using radiation therapy planning images.Materials and methodsFifty lung cancer patients from two institutions were analyzed, and a clinician contoured all five lung lobes [left upper lobe (LUL), left lower lobe (LLL), right upper lobe (RUL), right middle lobe (RML), and right lower lobe (RLL)] on the free-breathing computed tomography data set. The AI model used a residual 3D U-Net and trained using the expert lobe contours for forty patients. Validation was carried out by comparing expert lobe contours on ten patients against AI-based lobe contours using dice similarity coefficients (DSC).ResultsThe AI-based model showed good agreement with expert contours with overall DSC of 0.93 (range of 0.78–0.97). The DSC were 0.95 (0.97–0.91), 0.92 (0.96–0.85), 0.94 (0.97–0.87), 0.88 (0.93–0.78), and 0.94 (0.96–0.91), for the LUL, LLL, RUL, RML, and RLL, respectively.ConclusionsThis work presents a validation of AI-based lung lobe contours on free-breathing data and shows good agreement with expert contours. Introduction The American Cancer Society estimates that the total number of new cancer cases of the respiratory system will be approximately 256,000 for 2023 in the United States [1]. Radiation therapy remains an important option for the treatment of lung cancer. Pulmonary toxicity (PT) is a common treatment related toxicity and can include radiation pneumonitis and radiation fibrosis. Traditional methods of evaluating and mitigating pulmonary toxicity have relied on total lung dose metrics (mean lung dose, for example) or single lung dose metrics [2] (for example, ipsilateral mean lung dose). The lungs are divided into functional sub-units referred to as lung lobes. Lung lobes are separated from each other with tissue fissures. There are 3 lung lobes in the right lung and 2 lung lobes in the left lung. Data has shown that tumor location within the lung lobes may impact the risk of PT [2, 3], with inferior locations being associated with higher incidence of PT. Although tumor location presents strong associations with the induction of PT, the exact cause of this remains elusive. Vogelius and Bentzen [3] have postulated that the increased tumor motion, and treatment volume, may lead to increased PT. Furthermore, there has been growing interest in evaluating lung dose metrics on a lobar level. A recent study found that a higher mean dose to the lower lung lobes was a predictor of radiation pneumonitis [4]. The challenge of evaluating lobar dosimetry for lung cancer is that delineating lung lobes can be challenging and is time consuming. Artificial intelligence (AI) has been used in radiotherapy to automatically generate treatment planning contours in numerous sites with good agreement against clinically used contours [5, 6]. AI-based methods have been used for auto-segmentation7 and have the potential to reduce the time needed to contour lung lobes and to improve the accuracy of lung lobe contours. In the radiology domain, AI-based methods have been developed to accurately delineate the lobes of the lung [8], which in certain cases are required prerequisites for an accurate diagnosis. The challenges of generating AI-based lung lobe contours in radiotherapy applications is that the acquired images are different between radiology and radiation oncology. The radiology thoracic computed tomography (CT) scans are generally acquired in a breath-hold state while radiotherapy CT scans are acquired in a free breathing state and with different scan parameters (radiology scans typically have higher image resolution when compared to radiotherapy scans). As such, here has been limited research on generating AI-based lung lobe contours for radiotherapy planning scans [9]. Evaluating dosimetry and lung function at the lobar level has the potential to improve the ability to assess and predict pulmonary toxicity [10, 11]. Data on generating AI-based lobe contours on radiotherapy treatment planning CT scans are currently limited. The purpose of this study is to characterize and evaluate an AI-based lung lobe auto-contouring method using multi-institutional radiotherapy treatment planning CT images. Materials and methods Data from 50 patients who underwent thoracic radiotherapy treatment planning from two institutions (Ethical approval IRB# 21F.644) were used in this study. All patients underwent free-breathing and 4-dimensional computer tomography (4DCT) scans. 4DCT scans were acquired on either a Siemens (Munich, Germany) or GE Healthcare (Illinois, USA) scanner. The free breathing and 4DCT scans on the Siemens scanner were acquired using a gated lung protocol with 120 kVp, 3-mm thick slices, variable mAs, and the free breathing and 4DCT scans on the GE scanner were acquired using 2.5 mm thick slices, 120 kVp and variable mAs. In clinical practice, free breathing data sets are frequently used for dose calculations and were therefore used to develop the AI-based lobe segmentation model. Datasets were manually reviewed and excluded if significant motion artifacts were present including motion artifacts or lobar irregularities. Seven datasets were excluded due to motion artifacts. Lung lobes were contoured by a clinician (JL) on the free-breathing CT for all 50 patients, with a typical example shown in Figure 1. Of the 50 patient scans, 40 were used for AI model development and 10 were used for AI model validation. Physician contours for 40 patients were used to train a convolutional neural network (CNN), based on the RefineNet [12] architecture. RefineNet is a multi-path refinement network designed for high-resolution semantic image segmentation. This network focuses on capturing and refining image features at multiple scales through a series of modules that integrate coarse semantic information from deep layers with fine-grained details available in shallower layers. In our implementation, we used ResNet 50 [13] as the encoder backbone and adapted the final decoder layer to include 6 channels, one for the total lung and one for each lung lobe. The model’s parameters were optimized using the stochastic gradient descent (SGD) optimizer, with the dice similarity coefficient (DSC) loss guiding the training process. The trained model was deployed and used to generate lung lobes contours on the validation data in MIM Software (MIM Software Inc., Cleveland, Ohio). At the time of the study, the AI lung lobe model was not commercially available and only limited for research use. Future availability of the AI model will be available using commercial licensing with timing based upon the commercialization plans of the vendor. To evaluate the model’s performance and robustness, we implemented a five-fold cross-validation method with the forty scans used for model development. In this process, five different models were trained on subsets of 32 scans and evaluated on the 8 remaining scans within each fold. This approach allowed us to generate predicted lung lobe contours for all forty cases for evaluation. The model, trained on the first fold, was further validated using a separate set of 10 scans initially reserved for final validation. The DSC five-fold cross validation results are presented. The AI contours were generated on the 10 patient datasets used for independent validation and were compared to the manually expert drawn contours. The metrics used to compare the contours were DSC and maximum Hausdorff distance (HD). Results are provided as comparisons of the 5 lobes of the lung; the right upper lung (RUL), the right middle lung (RML), the right lower lung (RLL), the left upper lung (LUL) and the left lower lung (LLL), along with total DSC and HD values for all the lobes combined. Data is reported using descriptive statistics with means, ranges, and standard deviations. The presented ranges represent maximum and minimum values for the 10 patient dataset. Results The initial model developed in the five-fold cross-validation method displayed good agreement with the following DSC values: for the left lung, the LUL DSC of 0.92 ± 0.14 (mean ± standard deviation), the LLL DSC of 0.93 ± 0.07. For the right lung, the RUL DSC of 0.93 ± 0.04, the RML DSC of 0.86 ± 0.15, and the RLL of 0.93 ± 0.06. This internal validation was performed successfully, and acceptable for further validation with the final ten patients. The AI-based model showed overall good agreement with expert contours with the DSC for the overall data set of 0.93 ± 0.04 (range of 0.78–0.97), and HD of 38.2 mm ± 19.39 (range of 11.1–88.7). Lobe specific DSC’s and HD’s are provided in Tables 1 and 2. Figure 2A shows an example case where the contours of the left lung lobes are in good visual agreement with physician contours, even in the presence of a large mass. Figure 2B is another visual comparison between manual and AI based contours at various axial slices, focusing on the RML. Finally, Figure 2C shows a coronal view of the AI based contours, which anatomically agrees well with a typical lung anatomy. Discussion The AI-based model using radiation therapy planning image sets achieved an overall DSC of 0.93 ± 0.04, and HD of 38.2 mm ± 19.39. These results show that the developed AI-based lung lobe contouring model can produce lung lobe contours with average DSC higher than 0.90 when compared to clinician-drawn contours. Our data can be compared with previous work generating AI-based lung lobe and other organ at risk contours [14, 15]. Gu et al. used a 2D–3D hybrid CNN to develop AI-based lung lobe contours on CT images acquired for the treatment planning process. They reported DSC ranging from 0.90 to 0.96 [9]. Zhang et al. used a CNN to generate AI-based lung lobe contours on lung cancer screening low dose CT scans and reported DSC of 0.94 [14]. Khanna et al. trained a deep residual U-Net CNN using publicly available chest CT data and reported DSC of 0.98 [16]. Overall, the results from the current work are in line with previous studies showing that lung lobe contours can be generated with greater than 0.9 DSC agreement with clinically drawn lung lobe contours. The data from the current work can be further compared with AI-based organ at risk contours clinically validated in radiotherapy [5, 6]. Yang et al. report on the results of the thoracic radiation treatment planning auto-segmentation challenge and note high-performing auto-contours for the heart and lungs with DSCs ranging from 0.93 to 0.95 [5, 6]. The lung lobe contour DSC results presented in the current work are in line with previous clinical validation organ at risk results indicating that AI-generated lung lobes can produce clinically acceptable contours. Although the AI-based lung lobe contours achieved high DSC values, there were cases with sub-optimal agreement between the AI-based lung contours and the clinical contours. Sub-optimal agreement results were particularly present in the RML (Tab. 1 and 2). The reduced accuracy results in the RML can likely be attributed to higher difficulty in discerning the fissure between the RML and RUL. Figure 2B shows a difference in the AI-based contours against the clinical contours of the RML. This example shows that while AI-based lung lone contours produce excellent overall agreement with clinical lung lobe contours, examples exist where the contours are not in agreement, demonstrating that the AI-based lung lobe contours should be clinically reviewed. AI-based auto-contouring has reduced contouring treatment time and improved contour consistency in radiotherapy. The data in the current work extends the AI-based auto-contouring to a functional sub-unit level for an organ. Auto-contouring of lung lobes can enable the evaluation of dosimetry and function on a lung functional sub-unit level in a time frame suitable for busy clinical environments. An AI-based model for lung lobes was developed and validated using free-breathing thoracic CT scans that were acquired for treatment planning purposes. Previous AI-based lung lobe models have been typically developed using breath-hold images [8]. However, breath-hold images are rarely acquired in thoracic radiation oncology. It was therefore necessary to develop and validate an AI-based model on free-breathing images that are more frequently acquired for treatment planning in thoracic radiation oncology. Therefore, the impact of our work is that we developed and validated an AI-based model for lung lobes on free breathing scans such that the lung lobe contours can be generated during the radiation planning process. There are two drawbacks that arise due to the multi-institutional nature of the study. First, the free breathing scans were performed on different scanners, and used similar, but slightly different slice thicknesses (2.5 vs. 3mm). This could impact the quality of the scan, and the ability to discern the lung lobes. We believe the 2.5 mm and 3 mm slice thickness results presented in the current work represent the worst case scenario as radiation oncology CT scans are unlikely to have slice thickness < 3 mm. The current study did not evaluate the ability of the AI lung lobe algorithm using breath-hold scans (which can be acquired for radiation treatment planning). We hypothesize that the algorithm would perform equally well and likely better in breath-hold CTs due to decreased motion. Lastly, while fifty patients were used during this study, forty of them were used to create the model. An idealized study would have included a larger patient cohort. However, the data still showed promising results using a fifty-patient dataset. Conclusions The presented work developed and evaluated an AI-based lung lobe auto-contouring model. The AI-based lung lobe model showed good agreement against clinician drawn lung lobe contours with DSC for the overall data set of 0.93 ± 0.04 (range of 0.78–0.97). Our results demonstrate that AI-based auto-contouring can produce clinically acceptable lung lobe contours. Auto-contouring of lung lobes can enable the evaluation of dosimetry and function on a lung functional sub-unit level in a time frame suitable for busy clinical environments.