Goenka, Ajit H.

Abstract Purpose: Around 30% of PDAC less than 2-cm tend to go undetected on CT due to their subtle imaging signatures. Automated detection of PDAC using AI represents an opportunity to augment physician expertise and to improve outcomes through early detection of PDAC. Our purpose was to develop a 3D-CNN for fully automated detection of PDAC and to further evaluate the impact of inclusion of pancreas segmentation on the accuracy of this 3D-CNN. Methods: A Medical Imaging Data Readiness Scale (MIDaR) level A dataset (portal venous phase CTs, slice thickness ≤ 3.75 mm) of 466 treatment-naïve biopsy-proven PDAC and 1994 subjects with normal pancreas was created after exclusion of CTs with suboptimal image quality or biliary stents. Volumetric pancreas and tumor segmentations on CTs were done by two radiologists using 3D Slicer. A total of 370 CTs with PDAC and 370 CTs with normal pancreas were randomly selected for separate training and validation sets, and 396 CTs (96 CTs with PDAC and 300 CTs with normal pancreas) were utilized for testing. Two separate 3D-CNNs were trained. A three-stage bounding-box-only model (A): stage 1 was based on a UNET-like architecture and localized the pancreas on CT with a bounding box; stage 2 utilized an Inception ResNet architecture and classified each slice through the pancreas into PDAC vs. normal; and stage 3 utilized the output of stage 2 to generate final classification for a given CT. Conversely, a four-stage pancreas segmentation-based model (B) included stage 1 of model A followed by an additional stage of automated pancreas and tumor segmentation (stage 2), classification of each slice through the pancreas into PDAC vs. normal (stage 3) and, finally, generation of final classification score (stage 4) for a given CT. Area under the receiver operating characteristic curve (AUROC) of the two models were compared on the test set. Results: Mean (SD) PDAC diameter in the test set was 1.1 (0.43) cm. Model A (three-stage bounding-box-only) correctly classified 305 (77%) out of 396 CTs from the test set into PDAC vs. normal. It incorrectly classified 12/96 (12.5%) CTs with PDAC as normal and 79/300 (26%) normal CTs as PDAC. AUROC for model A was 0.85. Model B (four-stage pancreas segmentation-based) correctly classified 351 (88%) out of 396 CTs. It incorrectly classified 13/96 (13.5%) CTs with PDAC as normal and 32/300 (10.7%) normal CTs as PDAC. AUROC for model B was 0.94. AUROC for model B was significantly higher than model A (p<0.005). Conclusion: A 3D-CNN can detect small PDAC with high accuracy using automated localization of pancreas with a bounding box without relying on separate pancreas segmentation. Inclusion of an additional automated pancreas segmentation step reduced false positives with consequent incremental gain in the model’s accuracy. Prospective validation and subsequent integration of such models into clinical workflows has the potential to reduce inadvertent errors in detection of subtle or small PDAC on standard-of-care CT scans. Citation Format: Anurima Patra, Korfiatis Panagiotis, Garima Suman, Ananya Panda, Sushil Kumar Garg, Ajit Goenka. Automated detection of pancreatic ductal adenocarcinoma (PDAC) on CT scans using artificial intelligence (AI): Impact of inclusion of automated pancreas segmentation on the accuracy of 3D-convolutional neural network (CNN) [abstract]. In: Proceedings of the AACR Virtual Special Conference on Artificial Intelligence, Diagnosis, and Imaging; 2021 Jan 13-14. Philadelphia (PA): AACR; Clin Cancer Res 2021;27(5_Suppl):Abstract nr PO-084.

PurposeTo develop a two‐stage three‐dimensional (3D) convolutional neural networks (CNNs) for fully automated volumetric segmentation of pancreas on computed tomography (CT) and to further evaluate its performance in the context of intra‐reader and inter‐reader reliability at full dose and reduced radiation dose CTs on a public dataset.MethodsA dataset of 1994 abdomen CT scans (portal venous phase, slice thickness ≤ 3.75‐mm, multiple CT vendors) was curated by two radiologists (R1 and R2) to exclude cases with pancreatic pathology, suboptimal image quality, and image artifacts (n = 77). Remaining 1917 CTs were equally allocated between R1 and R2 for volumetric pancreas segmentation [ground truth (GT)]. This internal dataset was randomly divided into training (n = 1380), validation (n = 248), and test (n = 289) sets for the development of a two‐stage 3D CNN model based on a modified U‐net architecture for automated volumetric pancreas segmentation. Model’s performance for pancreas segmentation and the differences in model‐predicted pancreatic volumes vs GT volumes were compared on the test set. Subsequently, an external dataset from The Cancer Imaging Archive (TCIA) that had CT scans acquired at standard radiation dose and same scans reconstructed at a simulated 25% radiation dose was curated (n = 41). Volumetric pancreas segmentation was done on this TCIA dataset by R1 and R2 independently on the full dose and then at the reduced radiation dose CT images. Intra‐reader and inter‐reader reliability, model’s segmentation performance, and reliability between model‐predicted pancreatic volumes at full vs reduced dose were measured. Finally, model’s performance was tested on the benchmarking National Institute of Health (NIH)‐Pancreas CT (PCT) dataset.ResultsThree‐dimensional CNN had mean (SD) Dice similarity coefficient (DSC): 0.91 (0.03) and average Hausdorff distance of 0.15 (0.09) mm on the test set. Model’s performance was equivalent between males and females (P = 0.08) and across different CT slice thicknesses (P > 0.05) based on noninferiority statistical testing. There was no difference in model‐predicted and GT pancreatic volumes [mean predicted volume 99 cc (31cc); GT volume 101 cc (33 cc), P = 0.33]. Mean pancreatic volume difference was −2.7 cc (percent difference: −2.4% of GT volume) with excellent correlation between model‐predicted and GT volumes [concordance correlation coefficient (CCC)=0.97]. In the external TCIA dataset, the model had higher reliability than R1 and R2 on full vs reduced dose CT scans [model mean (SD) DSC: 0.96 (0.02), CCC = 0.995 vs R1 DSC: 0.83 (0.07), CCC = 0.89, and R2 DSC:0.87 (0.04), CCC = 0.97]. The DSC and volume concordance correlations for R1 vs R2 (inter‐reader reliability) were 0.85 (0.07), CCC = 0.90 at full dose and 0.83 (0.07), CCC = 0.96 at reduced dose datasets. There was good reliability between model and R1 at both full and reduced dose CT [full dose: DSC: 0.81 (0.07), CCC = 0.83 and reduced dose DSC:0.81 (0.08), CCC = 0.87]. Likewise, there was good reliability between model and R2 at both full and reduced dose CT [full dose: DSC: 0.84 (0.05), CCC = 0.89 and reduced dose DSC:0.83(0.06), CCC = 0.89]. There was no difference in model‐predicted and GT pancreatic volume in TCIA dataset (mean predicted volume 96 cc (33); GT pancreatic volume 89 cc (30), p = 0.31). Model had mean (SD) DSC: 0.89 (0.04) (minimum–maximum DSC: 0.79 −0.96) on the NIH‐PCT dataset.ConclusionA 3D CNN developed on the largest dataset of CTs is accurate for fully automated volumetric pancreas segmentation and is generalizable across a wide range of CT slice thicknesses, radiation dose, and patient gender. This 3D CNN offers a scalable tool to leverage biomarkers from pancreas morphometrics and radiomics for pancreatic diseases including for early pancreatic cancer detection.

Abstract Background We aim to assess the impact of a multidisciplinary inflammatory bowel disease (IBD) conference on the management of patients with complex IBD. Methods Data were collected during each conference from February 2017 through October 2018 with chart review performed at 6 months to determine if conference recommendations were successfully implemented. Results Eighty-five patients were discussed. The presenting diagnosis was changed by 11.8%. Recommendations for further testing, medical therapy, and surgery were successfully implemented in 77.1%, 98.1%, and 88.4%, respectively. Conclusions This study supports the role of IBD multidisciplinary conferences in the management of patients with complex IBD.

432 Background: In clinical trials and clinical practice, tumor response assessment with computed tomography (CT) defines critical end points in patients with metastatic disease treated with systemic agents. Methods to reduce errors and improve efficiency in tumor response assessment could improve patient care. Methods: Eleven readers from 10 different institutions independently categorized tumor response according to three different therapeutic response criteria using paired baseline and initial post-therapy CT studies from 20 randomly selected patients with metastatic renal cell carcinoma treated with sunitinib as part of a completed phase III multi-institutional study. Images were evaluated with a manual tumor response evaluation method (standard-of-care) and with computer-assisted response evaluation (CARE) that included stepwise guidance, interactive error-identification and correction methods, automated tumor metric extraction, calculations, response categorization, and data/image archival. A cross-over design, patient randomization, and two-week washout period were used to reduce recall bias. Comparative effectiveness metrics included error rate and mean patient evaluation time. Results: The standard-of-care method was on average associated with one or more errors in 30.5% (6.1/20) of patients while CARE had a 0.0% (0.0/20) error rate (p<0.001). The most common errors were related to data transfer and arithmetic calculation. In patients with errors, the median number of error types was 1 (range 1-3). Mean patient evaluation time with CARE was twice as fast as the standard-of-care method (6.4 vs. 13.1 minutes, p<0.001). Conclusions: Computer-assisted tumor response evaluation reduced errors and time of evaluation, indicating better overall effectiveness than manual tumor response evaluation methods that are the current standard-of-care.