Research Presentation Session: Imaging Informatics / Artificial Intelligence and Machine Learning
Author Block: M. Gross1, M. Spektor2, A. Jaffe2, A. S. Kücükkaya1, S. Iseke3, S. Haider4, M. Strazzabosco2, J. A. Onofrey2, J. Chapiro2; 1Berlin/DE, 2New Haven, CT/US, 3Rostock/DE, 4Munich/DE
Purpose or Learning Objective: Manual liver segmentation is time-consuming and expensive. In this study, a DCNN was trained using a novel framework that incorporates a small set of manually annotated imaging with a large set of automated segmentations to supplement segmentation performance.
Methods or Background: This study included 617 arterial-phase T1-weighted MR images of which 219 had corresponding manual liver segmentations. The 219 annotated images were split into 50/15/35% (n=109/33/77) training/validation/testing subsets, respectively. First, ten proportions (10-100%, increments of 10%) of manual liver segmentations from the training pool were used to train ten baseline (BL) DCNN models with identical 3D U-net architectures. The BL models were used to generate automated liver segmentations on the unlabeled dataset (n=398) and on the remaining portion of the training pool that was not in their respective training set. Second, these automated segmentations were combined with the manually annotated images into new training sets and used to train ten enhanced-training (ET) DCNNs. The dice similarity coefficient (DSC) was used to quantify segmentation performance and a Wilcoxon signed-rank test was used for comparisons.
Results or Findings: BL models trained with more manual segmentations substantially outperformed BL models trained with fewer segmentation data. ET models significantly outperformed their respective BL models (for all training sets including 10-50% of available manual segmentations: p<0.05). Some ET models significantly outperformed BL models trained with more manual segmentations (ET 30% vs BL 40%, ET 70% vs BL 90%, p<0.05).
Conclusion: This new training approach reduced the overfitting of models trained on smaller training sets and achieved satisfactory liver segmentation performance even with fewer expert annotations, outperforming models that were trained with more manual segmentations.
Limitations: One network architecture was used and a single contrast phase.
Ethics committee approval: This study was approved by the local IRB.
Funding for this study: This study was funded, NIH grant P30 KD034989.
Author Block: U. Fehrenbach1, S. Xin1, T. A. Auer1, H. Jann1, H. Amthauer1, D. Geisel1, T. Denecke2, B. Wiedenmann1, T. Penzkofer1; 1Berlin/DE, 2Leipzig/DE
Purpose or Learning Objective: Fast and exact quantification of hepatic metastasis is an unmet medical need in patients with secondary liver malignancies. We therefore present a deep-learning 3D-quantification model of neuroendocrine liver metastases (NELM) using gadoxetic-acid (Gd-EOB)-enhanced MRI.
Methods or Background: In 149 patients, manual segmentations of NELM and livers were used to train a neural network (278 Gd-EOB MRI scans). Clinical utility was evaluated in another 33 patients which were discussed in our multidisciplinary cancer conference (MCC) and received a Gd-EOB MRI both at baseline and as follow-up examination (n = 66). The model's measurements (NELM volume; hepatic tumour load (HTL)) with corresponding absolute (ΔabsNELM; ΔabsHTL) and relative changes (ΔrelNELM; ΔrelHTL) between baseline and follow-up were compared to MCC decisions of therapy response.
Results or Findings: Internal and external validation of the model’s accuracy showed a high overlap for NELM and livers (Matthew’s correlation coefficient (phi): 0.76/0.95 (internal), 0.86/0.96 (external)) with higher phi in larger NELM volume (phi= 0.80 vs. 0.71; p = 0.003). MCC decisions were significantly differentiated by all response variables (ΔabsNELM; ΔabsHTL; ΔrelNELM; ΔrelHTL) (p < 0.001). ΔrelNELM and ΔrelHTL showed optimal discrimination between therapy success or failure (AUC:1.000; p < 0.001).
Conclusion: The deep-learning based model shows high accuracy in 3D-quantification of NELM and HTL in Gd-EOB-MRI. The model’s measurements correlated well with the evaluation of therapeutic response of an expert MCC.
Limitations: The 3D assessment approach needs to be further evaluated in direct comparison to 2D measurements and its impact on clinical endpoints in larger cohorts. The ground truth of accuracy is based on manual segmentation of liver metastasis. Due to the sometimes pronounced, even small foci of liver metastases, manual segmentation is not perfect.
Ethics committee approval: Approved by the Institutional Review Board of Charité Berlin.
Funding for this study: No funding was received for this study.
Author Block: R. Park, S. S. Lee, Y. S. Sung, J. S. Yoon, H-I. Suk, H. J. Kim, S. H. Choi; Seoul/KR
Purpose or Learning Objective: To devise construct graft volume-to-weight conversion formula and to evaluate efficiency and accuracy of DLA-assisted CT volumetry in right lobe (RL) graft in a large cohort of LDLT.
Methods or Background: We retrospectively enrolled 581 RL donors and divided them into development and validation groups. The CT was analysed using DLA-assisted software. The graft volume-to-weight conversion formula was derived from the development group by linear regression. The agreement between estimated and measured graft weights and inter-reader agreement were assessed using CCC and 95% Bland-Altman LOA in the validation group. To assess factors influencing estimation error, multivariable linear regression was performed in the validation group.
Results or Findings: Segmentation correction was required in 28.6% cases with short correction time (mean, 12.8±33.6 seconds) and small change in volume (95% LOA, -3.0% to 3.0%). The total process time ranged from 1.3 to 8.0 minutes (mean, 1.8±0.6 minutes). The conversion formula was as follows: estimated graft weight (g)=206.3+0.653 x CT-measured graft volume (ml) (r=0.878, p<.001). Between estimated and measured graft weights, CCC was 0.834 and 95% LOA was -1.7%±17.1% (P=.002). Of the inter-reader agreement of RL volume, CCC was 0.998 and 95% LOA was 0.2%±1.8% (P=.069). The sex (coefficient;-3.47) and BMI (coefficient;-0.58) have a significant independent association with estimation error, while age, hepatosteatosis, CT interval and graft type did not.
Conclusion: We proposed a graft volume-to-weight conversion formula that would be useful in preoperative graft weight estimation. The DLA-assisted CT volumetry is a highly efficient method for preoperative graft weight estimation. The error margin of RL graft weight is within approximately 17% of graft weight.
Limitations: First, it was retrospective. Second, we evaluated only RL graft donors. Third, development and validation groups were enrolled in the same institution.
Ethics committee approval: IRB waived the requirement for informed consent.
Funding for this study: This research was supported by a National Research Foundation of Korea (NRF) grant, funded by the Korean government (MSIT) (2020R1F1A1048826).
Author Block: M. Ghelichoghli1, s. m. bagheri2, A. Akhavan1, V. Ashkani Chenarlogh1, N. Sirjani1, I. Shiri3, A. Shabanzadeh1; 1Karaj/IR, 2tehran/IR, 3Geneva/CH
Purpose or Learning Objective: In this study, we have proposed and validated a fully automatic approach for the quantification of fatty liver disease using ultrasound images based on hepato-renal-index (HRI) calculation. The procedure includes segmentation of kidney and liver, detection of an ROI in the renal parenchyma region and liver at the same depth, and HRI calculation.
Methods or Background: We proposed a highly accurate and fast convolutional neural network, named Fast-Unet, for the segmentation of kidneys and liver. The main superiority of Fast-Unet model is low response time, which is appropriate in ultrasound image analysis that needs on-site measurement by radiologists. We used a superpixel algorithm to find the lowest variance region in parenchyma as the renal ROI. At the next step, all pixels with the same depth as renal ROI centrum were found using the intersection of two borders of the convex probe sector. This step is conducted because if the renal and liver ROIs were not in the same depth resulting HRI is not accurate due to the ultrasound depth attenuation effect. Finally, an ROI in the liver with the same depth was found and HRI was calculated.
Results or Findings: The train-test dataset contained 752 ultrasound images. The Dice and Jaccard coefficients were used to evaluate the segmentation step, and 94% and 89% for the kidney and 97% and 91% for the liver were achieved respectively. The predicted HRI values were also validated with a radiologist's report using the root-mean-square-error (RMSE) metric and 0.04 was achieved.
Conclusion: Automation of HRI calculation speeds up the fatty liver diagnosis and helps novice radiologists to interpret ultrasound images more accurately.
Limitations: There is no limitation in this study.
Ethics committee approval: Med Fanavaran Plus co. ethics committee approved this study.
Funding for this study: Med Fanavaran Plus Co. funded this study.
Author Block: W. Li1, Y. Zhu1, G. Zhao1, X. Chen1, X. Zhao2, Q. Xie1, F. Yan1; 1Shanghai/CN, 2Guangdong/CN
Purpose or Learning Objective: To evaluate the performance of fully automated deep learning (DL) algorithm for staging hepatic fibrosis and distinguishing fibrosis from normal people based on MR images with or without non-image information.
Methods or Background: 500 patients were retrospectively enrolled from two hospitals. Model DL were built using delay phase MR images to assess fibrosis stages. In addition, different models of model DL combined with non-image information including biomarkers (APRI and FIB-4), virus status (hepatitis B and C virus tests), and MR information (manufactures and static magnetic field).The AUROCs were compared between different models using Delong test, the sensitivity and specificity of both model DL and model Full (model DL combined with all non-image information) were compared with experienced radiologists and biomarkers using McNemar’s test.
Results or Findings: In the test set, the AUROC (with 95% confidence intervals) values of model Full for diagnosing fibrosis stages F0-4, F1-4, F2-4, F3-4 and F4 were 0.99 (0.94-1.00), 0.98 (0.93-0.99), 0.90 (0.83-0.95), 0.81 (0.73-0.88) and 0.84 (0.76-0.90), respectively, which outperformed model DL on diagnosing F0-4 and F1-4.Compared with the radiologists, model Full showed better specificity for fibrosis stage F0-4, better sensitivity for the other four classification tasks. While compared with biomarkers, both model DL and model Full showed significantly higher specificity in staging F3-4 and F4.
Conclusion: DL using MR images with or without non-image data provides a promising non-invasive assessment tool for the staging liver fibrosis, and for distinguishing liver fibrosis patients from normal people, with a performance superior to experienced radiologists and biomarkers.
Limitations: The sample among fibrosis stage was unbalanced.
Ethics committee approval: Ruijin Hospital affiliated to Shanghai JiaoTong University School of Medicine
Funding for this study: National Natural Science Foundation of China (grant numbers 81401406) and Innovative research team of high-level local universities in Shanghai.
Author Block: m. Taghavi, F. Staal, M. Maas, S. H. Benson, R. G. H. Beets-Tan; Amsterdam/NL
Purpose or Learning Objective: For solitary small (< 3 cm) tumors, ablation of liver metastases is now part of international guidelines. However, a significant number of patients experience regrowth due to insufficient treatment or inadequate ablation of tumour margins. Local tumour progression (LTP): defined as regrowth /recurrent disease after ablation is a threat to overall survival. The aim of this study was to create a deep learning model using CT images obtained at baseline and directly after treatment in order to predict which patients would later experience regrowth.
Methods or Background: For this study, we retrospectively included 79 patients (120 lesions) with colorectal liver metastasis (CRLM) who were treated by thermal ablation consisting of either radiofrequency ablation (RFA), or microwave ablation (MWA) for liver metastases (LM). Exclusion criteria were based on the ESMO guidelines. The pre- and post-treatment scans were used as input to a multi-channel Convolutional Neural Network (CNN). The manual lesion delineation was used to identify a 3D region of interest (RoI) around each lesion. We employed transfer learning in order to train a deep learning model for the dataset in question. A 19-layer CNN from the Visual Geometry Group (VGG-19) was found to perform best.
Results or Findings: The area under the receiver operating characteristic curve (AUC) was found to be 0.72, 95% confidence interval: (0.64, 0.79).
Conclusion: We have demonstrated that it is possible to use transfer learning together with CNN models in order to predict tumour progression and also demonstrated that it is possible to employ state-of-the-art methods to avoid overfitting.
Limitations: Small cohort size of 79 patients, therefore impacting the size of the AUC confidence interval.
Ethics committee approval: The informed consent requirement was waived by the Institutional Review Board due to the retrospective nature of the study.
Funding for this study: Not applicable.
Author Block: M. Ligero, E. Delgado, J. Hernando, A. Garcia-Alvarez, X. Merino Casabiel, M. Escobar, J. Capdevila, R. Perez Lopez; Barcelona/ES
Purpose or Learning Objective: To define and validate CT-based radiomics phenotypes associating with response to anti-angiogenic treatment in patients with gastroenteropancretic neuroendocrine tumors(GEP-NET). To investigate if multi-phase radiomics model or the combination of radiomics with clinical data improves response prediction.
Methods or Background: A predictive CT-based radiomics signature was developed in 57 patients included in the TALENT phase II prospective trial of lenvatinib in advanced GEP-NET from October 2015 to August 2020. Radiomics features were extracted from all liver lesions at pre-treatment CT. Features were selected using minimum redundancy maximum relevance(mRMR) and combined in a logistic regression for predicting clinical benefit (progression free survival[PFS]>15 months). A multiphase model including arterial and portal acquisitions was developed. Models were validated internally and tested in an external cohort of 26 patients treated with the VEGFR1-3 inhibitor sunitinib. A regression model was used to combine radiomics and clinical variables. Model interpretability plots were also developed.
Results or Findings: In the training and validation set, the model associated with response (area under the curve [AUC] 0.76 and 0.69, respectively). In the test set the model associated with response (AUC = 0.68). The multi-phase didn’t improve the prediction capacity (AUC = 0.59). The model combining radiomics and clinical variables slightly improved the capacity for predicting response in the three cohorts (AUC = 0.78, 0.75 and 0.67, respectively). Interpretability plots showed that patients with high radiomics-score presented more spherical and hypervascularised lesions.
Conclusion: Single-phase radiomics associates with response to anti-angiogenics based on the quantification of hypervascularisation and tumour shape.
Limitations: Further testing populations are needed to validate the prediction capacity of the model.
Ethics committee approval: The institutional review board approved this retrospective study. Need for informed consent for the computational analysis of the images was waived.
Funding for this study: The TALENT clinical trial was funded by Eisai.