Search

Purpose/objective(s): Identifying suspicious cancer metastasized lymph nodes (GTV is an essential task in esophageal cancer radiotherapy. Manual delineation requires high-level sophisticated clinical reasoning in multi-modality imaging, which is challenging and suffers from high subjectivity and inter-observer variability. We propose an automated esophageal GTV_LN detection method via integrating a distance-based decision stratification into a multi-task deep network, which is able to effectively learn the GTV_LN oncology features. This is inspired by the fact that lymph node involvements often follow certain distance patterns from the primary tumors.Materials/methods: We collected and curated 2 esophageal cancer datasets with PET and RTCT images. The 1st dataset contained 141 patients and 651 annotated GTV_LN for training and evaluating our deep network (60%, 10% and 30% for training, validation and testing), while the 2nd dataset contained 10 patients aiming for the multi-user agreement analysis. A distance-based decision stratification is designed to divide GTV_LN into "tumor-proximal" and "tumor-distal" categories and a multi-branch deep network is proposed to solve each of them. The multi-branch network has a shared encoder and separate decoders to detect and segment two GTV_LN subcategories, respectively. The distance-based decision stratification enables the automated multi-branch method to specialize in GTV_LN oncology features of each subcategory. We use both PET and RTCT as the network inputs to exploit the complementary information in each imaging modality.Results: In the testing set of the 1st esophageal cancer dataset (34 patients with 138 GTV_LNs), our method achieves a sensitivity of 73.8% and 79.1% at 6 and 12 false positives (FPs) per patient as compared to 72.0% and 75.9% by MULAN, the state-of-the-art deep learning based universal lesion detection algorithm. On the 2nd dataset containing 10 patients, 3 radiation oncologists annotated 53 GTV_LNs in total, among which 42 GTV_LNs were agreed with all 3 physicians leading to 79.2% multi-user agreement. Our method achieves a sensitivity of 77.4% at 6 FPs per patient on the 2nd dataset. This shows that the sensitivity achieved by our automated deep learning approach can be close to the multi-user agreement.Conclusion: We proposed a multi-task deep learning method with distance-based decision stratification to effectively detect and segment the GTV_LN in esophageal cancer radiotherapy. It significantly improves on previous state-of-the-art and the achieved performance has potential clinical values considering the large multi-user variability in this challenging task. [ABSTRACT FROM AUTHOR]

Organs at risk (OARs) segmentation is an essential process in head and neck (H&N) cancer radiotherapy. We have reported high automated segmentation geometric accuracy of Stratified Learning and Neural Architecture Search method in terms of Dice Score (DSC) from our previous study. In this study, we would evaluate the dosimetric influence of our automated approach before integrating this method into the clinical workflow. To measure the dosimetric effects brought by the OARs' variance, the intensity-modulated radiotherapy (IMRT) dose plans of 10 head and neck cancer patients were replanned using the original tumor target volumes and three substitute OAR contours permutations (deep learning generated stratified organs at risk segmentation (SOARS), SOARS revised by physician (SOARS-revised), and OAR delineated from scratch by physician (human reader)). We further examined the clinical dosimetric accuracy and the clinical reference OAR contours were overlaid on top of each replanned dose grid to evaluate the dosimetric differences. After replanning, SOARS and SOARS-revised contours have slightly smaller Diff (max dose) as compared to human reader contours (3.4%, 3.5% vs. 4.1%). For the Diff (mean dose), human reader, SOARS, SOARS-revised achieves similar results, i.e., 5.3%, 5.0%, and 5.0%, respectively. However, more OARs from the human reader have dose variations larger than 10% or 20% as compared to SOARS and SOARS-revised. Overall, our results indicate that using OAR contours from human reader, SOARS, and SOARS-revised lead to generally comparable dose accuracy in clinical practice. SOAR-related OAR contours have fewer OARs with dose error larger than 10% or 20%. This study further validates the clinical applicability of a deep learning based automated H&N OAR segmentation method by comparing dosimetry of plans using OAR contours generated automatically and by a human reader to the gold standard contours. The dose variations calculated after planning on automated segmentation contours are less than 5%. Our proposed automated H&N OAR segmentation method not only achieves high geometric accuracy but also helps deliver treatment beams with little variances. [ABSTRACT FROM AUTHOR]

Supervised deep learning based automated delineating studies often rely on human experts for providing gold-standard contours and evaluating the contours delineated by deep learning models. Hence, the degree of intra-observer variation may affect the quality of deep learning-based delineating research. This study aims to evaluate the intra-observer variation in the contour reviewing process. We used 30 testing patients from an institution and enrolled a senior radiation oncologist who has initially delineated the contours of head and neck OARs of these patients as gold-standard. The testing treatment plan of each patient is composed of 13 head and neck OARs randomly selected organs from three sources: gold-standard, deep learning generated OAR contours and delineation by another radiation oncologist. The real contouring source for each OAR was kept unknown to this physician, and the physician had to decide if a particular OAR on each patient's treatment plan requires revision or not. The physician's decision on each OAR presented was then recorded and analyzed. 15% gold-standard contours required further editing, reflecting the intra-observer variation in deciding whether the revision of an OAR is needed. For deep learning generated contours, 43% requires revision, which is slightly higher than that in the earlier unblind assessment by this physician, where 37% deep learning generated contours required revision among the 13 OARs. However, since the number of deep learning generated contours requiring revision from two times study is close, it indicates that our intra-observer variation is within a small range. Moreover, compared with deep learning generated contours, a noticeable higher number of another human reader's contours requires revision (55% vs 43% of deep learning generated contours), reflecting that deep learning generated contours' quality is generally better than the human reader's in the blind assessment. This study has demonstrated that inherent intra-observer variation exists in a deep learning based automated delineating study. The observer disapproves 15% of own contouring in the blind reviewing experiment. Deep learning generated OAR contours remain comparable editing rate in the blind process and are of better quality than another human reader's manual contours. [ABSTRACT FROM AUTHOR]