Research Presentation Session: Abdominal Viscera & GI Tract

RPS 2401 - Artificial intelligence in abdominal imaging: current directions

March 3, 11:30 - 12:30 CET

7 min
Risk stratification of gallbladder masses by ML‑based ultrasound radiomics models: a prospective and multi‑institutional study
Chong-Ke Zhao, Shanghai / China
Author Block: C-K. Zhao; Shanghai/CN
Purpose: This study aimed to evaluate the diagnostic performance of machine learning (ML)–based ultrasound (US) radiomics models for risk stratification of gallbladder (GB) masses.
Methods or Background: We prospectively examined 640 pathologically confirmed GB masses obtained from 640 patients between August 2019 and October 2022 at four institutions. Radiomics features were extracted from greyscale US images and germane features were selected. Subsequently, 11 ML algorithms were separately used with the selected features to construct optimum US radiomics models for risk stratification of the GB masses. Furthermore, we compared the diagnostic performance of these models with the conventional US and contrast-enhanced US (CEUS) models.
Results or Findings: The optimal XGBoost-based US radiomics model for discriminating neoplastic from non-neoplastic GB lesions showed higher diagnostic performance in terms of areas under the curves (AUCs) than the conventional US model (0.822–0.853 vs. 0.642–0.706, p < 0.05) and potentially decreased unnecessary cholecystectomy rate in a speculative comparison with performing cholecystectomy for lesions sized over 10 mm (2.7–13.8% vs. 53.6–64.9%, p < 0.05) in the validation and test sets. The AUCs of the XGBoost-based US radiomics model for discriminating carcinomas from benign GB lesions were higher than the conventional US model (0.904–0.979 vs. 0.706–0.766, p < 0.05). The XGBoost-US radiomics model performed better than the CEUS model in discriminating GB carcinomas (AUC: 0.995 vs. 0.902, p = 0.011).
Conclusion: The proposed ML-based US radiomics models possess the potential capacity for risk stratification of GB masses and may reduce the unnecessary cholecystectomy rate and use of CEUS.
Limitations: First, the sample size was relatively limited. Second, since surgery-confirmed pathology was used as the gold standard, GB lesions sized less than 6 mm were not included, which are commonly not recommended for cholecystectomy.
Funding for this study: This work was supported in part by the National Natural Science Foundation of China (Grant 82202174).
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: This study was approved by the ethics committee of the institution (No: SHSYIEC–4.1/21–263/01; 2022-187R).
7 min
Development and validation of comprehensive nomogram based on imaging features and MRI radiomics to predict microvascular invasion and overall survival in patients with intrahepatic cholangiocarcinoma
Gengyun Miao, Shanghai / China
Author Block: G. Miao, X. Qian, Y. Zhang, C. Yang, M. Zeng; Shanghai/CN
Purpose: Microvascular invasion (MVI) is a predictor of poor prognosis in intrahepatic cholangiocarcinoma (ICC). The aim of this study was to establish a comprehensive model based on MR radiomics for MVI status stratification and overall survival prediction in ICC patients preoperatively.
Methods or Background: A total of 249 ICC patients were randomized into training and validation cohorts (174:75), and a time-independent test cohort with 47 ICC patients was enrolled. Independent clinical and imaging predictors were identified by univariate and multivariate logistic regression analyses. The radiomic model was based on the robust radiomic features extracted by a logistic regression classifier and the least absolute shrinkage and selection operator algorithm. The imaging-radiomics (IR) model integrated the independent predictors and robust radiomics features. The predictive efficacy of the models was evaluated by receiver operating characteristic curves, calibration curves and decision curves. Multivariate Cox analysis identified the independent risk factors for overall survival, Kaplan‒Meier curves were plotted, and a nomogram was used to visualize the predictive model.
Results or Findings: The imaging model comprised tumour size and intrahepatic duct dilatation. The radiomics model comprises 25 stable radiomics features. The IR model shows desirable performance (AUCtraining= 0.890, AUCvalidation= 0.885 and AUCtest= 0.815). The calibration curve and decision curve validate the clinical utility. Overall survival predicted by histological and IR model-predicted MVI groups exhibited similar predictive efficacy.
Conclusion: The IR model and nomogram based on IR model-predicted MVI status may be a potential tool in MVI status stratification and overall survival prediction of ICC patients preoperatively.
Limitations: The models are based on retrospectively collected data from a single institution.
Funding for this study: This work was supported by 1. Shanghai Municipal Health Commission (Grant number 202240152); 2. National Natural Science Foundation of China (Grant number 82171897); 3. Shanghai Municipal Key Clinical Specialty (Grant number shslczdzk03202); 4. Clinical Research Plan of SHDC (Grant number SHDC2020CR1029B).
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Approval for this retrospective study was granted by the Ethics Committee of our Hospital.
7 min
Automatic detection and segmentation of IPMN pancreatic cysts in MRI with a multi-sequence cascaded deep learning pipeline
Leo Joskowicz, Jerusalem / Israel
Author Block: N. Mazor, N. Lev-Cohain, R. Lederman, J. Sosna, L. Joskowicz; Jerusalem/IL
Purpose: Radiological detection and follow-up of IPMN pancreatic cysts in multi-sequence MRI studies is required to assess their malignancy potential. The evaluation requires expertise and is not automated. This study evaluates a novel multi-sequence cascaded deep learning pipeline for the detection and segmentation of IPMN pancreatic cysts in abdominal MRI.
Methods or Background: The pipeline consists of three steps: (1) pancreas Region of Interest (ROI) segmentation in the axial MRI TSE; (2) transfer and masking of the computed pancreas ROI to the coronal MRI MRCP; (3) detection and segmentation of cysts in the masked MRCP. Both steps 1 and 3 use 3D U-Net models with Hard Negative Patch Mining, a new technique for class imbalance correction and reduction of false positives. The pipeline was evaluated on 158 MRI patient studies of patients with pancreatic cysts undergoing follow-up. The training/validation/testing sets split was 118/17/23. Ground truth segmentations of a total of 840 cysts were manually obtained by an expert radiologist: Six hundred and nineteen cysts were >5 mm, 221 cysts were >10 mm with a mean number of cysts/scan of 5.3±2.6 and mean cyst diameter (volume) of 7.4 mm (0.91cc). The computed test set results were then compared to their respective manual ground truth delineations.
Results or Findings: The pipeline achieved mean recall of 0.80±0.19 and 0.99±0.05, precision of 0.75±0.26, and 0.95±0.16, and dice score of 0.80±0.19 and and 0.81±0.11 for pancreatic cysts of diameter > 5 mm and > 10mm respectively, which is the clinically relevant endpoint. The cyst inclusion in the pancreas ROI Recall is 0.94±0.22 and 0.98±0.07, respectively.
Conclusion: Automatic pancreatic cyst detection and segmentation in multi-sequence MRI may provide an accurate and reliable method for precise disease evaluation and save time.
Limitations: The major limitation of this study was this was of single observer annotation, one centre.
Funding for this study: No funding was obtained for this study.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: No ethical approval was obtained for this study.
7 min
Delta CT-radiomics derived response prediction in advanced pancreatic ductal adenocarcinoma
Felix N. Harder, Munich / Germany
Author Block: F. N. Harder1, G. O'Kane2, E. Salinas-Miranda2, K. Lajkosz2, A. Farooq2, S. Gallinger2, J. Knox2, M. Haider2; 1Munich/DE, 2Toronto, ON/CA
Purpose: The aim of this study was to evaluate multi-time point kinetic delta radiomics for response and overall survival (OS) prediction in advanced PDAC.
Methods or Background: One hundred and fifty-seven patients with advanced PDAC (108/157 with synchronous liver metastases) were retrospectively enrolled serving as the training cohort. Twenty-eight patients with metastatic PDAC from a second prospective study served as an external validation cohort. All patients received mFOLFIRINOX or gemcitabine/nab-paclitaxel as first-line chemotherapy. Baseline CT-models and delta radiomics models reflecting the kinetic between baseline and first follow-up CT were build based on size-related, non-size and combined size and non-size features from the primary tumour and largest liver metastasis to predict progression under chemotherapy based on RECIST1.1 and OS at 9 months in the overall cohort and liver metastases subgroup. Baseline and delta radiomics models were compared against each other and established Moffitt RNA-signature.
Results or Findings: Non-size and combined delta-radiomics models significantly discriminated between responders (complete/partial response and stable disease) and non-responders (progressive disease) in the training cohort and external validation cohort (AUC 0.714-0.873, p = < 0.001-0.01) outperforming baseline-only models (AUC 0.55-0.645) and Moffitt RNA signature (0.551-0.675). Radiomics models represented an independent survival predictor at 9 months in the training and the external validation cohort, with non-size models yielding the highest AUC in the training cohort, yet not significantly outperforming Moffitt RNA signature (training: 0.726 vs 0.588, p = 0.16; validation: 0.713 vs. 0.567, p = 0.35).
Conclusion: Delta-radiomics models outperformed baseline-models and Moffitt RNA signature as predictive biomarkers for response and OS prediction in advanced PDAC. In particular non-size feature models from the primary tumour and the largest liver metastasis provided additive value.
Limitations: Although including the largest cohort for radiomics-based response prediction in advanced PDAC so far, further studies need to validate the herein-found results.
Funding for this study: Funding for this study was provided by the Ontario Institute of Cancer Research, The Sinai Health Foundation and University Medical Imaging Toronto. F.N.H. received funding from Deutsche Forschungsgemeinschaft, HA 9949/1-1.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: This study was approved by our institutional research ethics board and informed consent was obtained.
7 min
Radiomic feature profiles to define treatment response in rectal cancer: not just a tumour matter
Ana Marhuenda, Valencia / Spain
Author Block: A. Marhuenda1, A. Nogue1, M. Domingo Pomar1, I. Machado1, R. Garcia Figueiras2, F. Bellvis1, A. Fuster Matanzo1, A. Jimenez-Pastor1, A. Alberich-Bayarri1; 1Valencia/ES, 2Santiago de Compostela/ES
Purpose: In rectal cancer (RC), predicting response to specific treatments is essential for defining appropriate therapeutic strategies. Moreover, the involvement of peritumoral regions and/or the presence of certain histopathological conditions are associated with poor outcomes. The project aimed to assess whether radiomics may help stratify patients at baseline based on treatment response and involvement of peritumoural regions.
Methods or Background: A retrospective, single-centre study was conducted. Baseline T2W MRIs of RC patients receiving neoadjuvant treatment (NT) and surgery were included. Manual delineation of seven labels was performed: tumour, extramural venous invasion (EMVI), tumoural deposits (TD), lymph nodes (LN) including intra- and extra- mesorectum nodes, peritumoural wall (PW), mesorectum fat, and presacral space (PS). Radiomic features were extracted for each segmentation, and four models were evaluated: tumour [model 1], tumour surroundings (EMVI, TD, LN, and mesorectum) [model 2], model 1 plus model 2 [model 3], and model 3 with PW and PS [model 4]. Univariate and multivariate (logistic regression) analyses were performed.
Results or Findings: A total of 50 RC patients who received any type of NT and underwent surgery were included. In the univariate analysis, the greatest differences between responders and non-responders were found in model 4. Statistical differences (p < 0.05) were noted in four radiomic features—Kurtosis, GLRLM Run Entropy, NGTDM_Busyness and NGTDM_Strength. In the multivariate analysis, model 4 outperformed the other models, with an AUC of 0.787.
Conclusion: Radiomic features could assist oncologists in therapeutic decision-making by predicting treatment responses. Segmentations including tumour and peritumoral regions provide more solid results. This highlights the relevance of a more holistic approach that would simplify segmentation´s tasks. Further studies with larger sample sizes are required.
Limitations: The limitations of the study are basically focused on the reduced number of patients.
Funding for this study: No funding was provided for this study.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: The study was approved by 2018-27.
7 min
Effect of artificial intelligence-aided differentiation of benign and premalignant colorectal polyps as a second reader at CT colonography
Sergio Grosu, Munich / Germany
Author Block: S. Grosu, M. P. Fabritius, M. Winkelmann, S. Maurus, A. Graser, J. Ricke, P. M. Kazmierczak, M. Ingrisch, P. Wesp; Munich/DE
Purpose: Premalignant adenomatous colorectal polyps require endoscopic resection, as opposed to benign hyperplastic colorectal polyps. The aim of this study was to evaluate the effect of artificial intelligence (AI)-assisted differentiation of benign and premalignant colorectal polyps as a second reader for general radiologists at CT colonography.
Methods or Background: CT colonography images with colorectal polyps of all sizes and morphologies were retrospectively evaluated by three independent board-certified radiologists with moderate experience in CT colonography. The readers’ task was to decide whether the depicted polyps required endoscopic resection. After a primary unassisted read, a second read with access to the classification of a radiomics-based random forest AI model labelling each polyp as “adenomatous” or “hyperplastic” was performed. No polyp used for training the AI model was included in this study. The performance of the unassisted and AI-assisted reading was evaluated using polyp histopathology as the reference standard.
Results or Findings: Seventy-seven polyps in 59 patients comprising 118 polyp image series (47% supine position, 53% prone position) were evaluated unassisted and AI-assisted by three radiologists, resulting in a total of 708 readings (subsequent polypectomy: yes or no). Compared with unassisted reading, the AI-assisted reading had a significantly higher accuracy (76% +/- 1% vs. 84% +/- 1%, p < 0.001), sensitivity (76% +/- 2% vs. 85% +/- 0%, p < 0.001), and specificity (75% +/- 1% vs. 81% +/- 2%, p < 0.001) in selecting polyps eligible for polypectomy.
Conclusion: In this proof-of-concept study, AI-based characterisation of colorectal polyps at CT colonography as a second reader enabled a more precise selection of polyps eligible for subsequent endoscopic resection.
Limitations: The limitation of this study is the rather small sample size.
Funding for this study: No funding was received for this study.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: This study was approved by the ethics committe of the Ludwig-Maximilians University, Munich (18-401).
7 min
A radiomics model for the preoperative prediction of lymph node metastasis in colorectal carcinoma
Alba Salgado-Parente, Madrid / Spain
Author Block: A. Salgado-Parente1, L. González Campo1, A. M. Vera Carmona1, A. Martínez Caballero1, I. De Vicente Bernal1, L. D. Juez Saez1, P. Abadía Barnó1, A. Torrado-Carvajal2, J. Blazquez Sanchez1; 1Madrid/ES, 2Mostoles/ES
Purpose: The aim of this study was to develop and validate a radiomics nomogram for the preoperative prediction of lymph node (LN) metastasis in colorectal carcinoma (CRC).
Methods or Background: The prediction model was developed in a primary cohort that consisted of 110 patients with clinicopathologically confirmed CRC with data collected from January 2013 to September 2017 in a single institution. The patients were divided into a training set (n = 88) and a validation set (n = 22) with statistically comparable demographic features. Radiomics features of the primary tumor and lymph node were extracted from portal venous phase CT images of each patient. Mutual information was used on the whole dataset for feature selection. The following models were trained: Random Forest, Logistic Regression, Naive Bayes, Gaussian Process, Support Vector Machine, MultiLayer Perceptron, K-Nearest Neighbors, Gradient Boosting, Neural Network. A crossvalidated fine-tuning of the hyperparameters was performed on each model to enhance the overall performance. A majority voting approach was followed assessing the different combinations of the individual classifiers. The combinations with the best F1 score performance were then selected to present the results.
Results or Findings: The radiomics signature demonstrated favorable discriminatory ability in predicting lymph node involvement based on both tumor segmentation (ROC AUC 0.88, sensitivity 1, specificity 0.73) and lymph node segmentation (ROC AUC 0.94, sensitivity 1, specificity 0.83). Accuracy was highly discriminative with values of 0.86 for tumor-based segmentations and 0.91 for lymph node-based segmentations.
Conclusion: The CT-based radiomics nomogram has the potential to be used as a non-invasive tool for individualised preoperative prediction of LN metastasis in CRC. External validation is further required prior to clinical implementation.
Limitations: The major limitations were that this was a single-centre and retrospective analysis.
Funding for this study: No funding was received for this study.
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information: Not applicable for this study.
7 min
Development and validation of CT-based radiomics deep-learning signatures to preoperatively predict lymph node metastasis in non-functional pancreatic neuroendocrine tumour: a multi-cohort study
Wei Tang, Shanghai / China
Author Block: W. Tang; Shanghai/CN
Purpose: Lymph node status is an important factor for the patients with non-functional pancreatic neuroendocrine tumours (NF-PanNETs) with respect to surgical methods, prognosis, recurrence. Our aim is to develop and validate a combination model based on contrast-enhanced CT images to preoperatively predict the lymph node metastasis (LNM) in NF-PanNETs.
Methods or Background: Retrospective data were gathered for 320 patients with NF-PanNETs who underwent curative pancreatic resection and CT imaging at two institutions (Centre 1, n= 236 and Centre 2, n=84) between January 2010 and March 2022. RDPs (Radiomics deep learning signature) were developed based on ten machine-learning techniques. These signatures were integrated with the clinicopathological factors into a nomogram for clinical applications. The evaluation of the model's performance was conducted through the metrics of the area under the curve (AUC).
Results or Findings: The RDP signature showed excellent performance in both centres with a high AUC for predicting LNM and DFS in centre 1 (AUC, 0.88; 95% CI: 0.84, 0.92; DFS, p <.05) and centre 2 (AUC, 0.91; 95% CI: 0.85, 0.97; DFS, p <.05). The clinical factors of vascular invasion, perineural invasion, and tumour grade were associated with LNM (p <.05). The combination nomogram showed better prediction capability for LNM (AUC, 0.93; 95% CI: 0.89, 0.96). Notably, our model maintained a satisfactory predictive ability for tumours at the 2-cm threshold, demonstrating its effectiveness across different tumour sizes in centre 1 (≤2 cm: AUC, 0.90 and >2 cm: AUC, 0.86) and centre 2 (≤2 cm: AUC, 0.93 and >2 cm: AUC, 0.91).
Conclusion: Our RDPs may have the potential to preoperatively predict LNM in NF-PanNETs, address the insufficiency of clinical guidelines concerning the 2-cm threshold for tumour lymph node dissection, and provide precise therapeutic strategies.
Limitations: Small sample size was a limitation of this study.
Funding for this study: This work was supported by Project of Shanghai Municipal Health Commission (202340123) and The Rare Tumour Research Special Project of the National Natural Science Foundation of China (82141104).
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: This two-centre retrospective study received approval from the Research Ethics Committee of the Institutional Review Boards from all participating hospitals, and the need for informed consent was exempted.

This session will not be streamed, nor will it be available on-demand!