RPS 416 - Radiomics and artificial intelligence

Author Block: A. Crombé, C. Lucchesi, F. Bertolo, M. Kind, A. Michot, R. Perret, F. Le Loarer, A. Bourdon, A. Italiano; Bordeaux/FR
Purpose: The purpose of this study was to identify subgroups of soft-tissue sarcoma (STS) patients using handcrafted and deep radiomics, to understand them, and investigate their impact on metastatic relapse-free survival (MFS).
Methods or Background: We included all consecutive adults with newly diagnosed locally-advanced STS managed at our sarcoma centre between 2008 and 2020, with contrast-enhanced (CE) baseline MRI. After MRI post-processing, segmentation and reproducibility assessment, 175 handcrafted radiomics features (h-RFs) from T1-weighted imaging (WI), T2-WI and fat-suppressed CE-T1-WI were calculated. Convolutional autoencoder neural network (CAE) and half-supervised CAE (HSCAE) were trained in repeated cross-validation on CE-T1-WI from one training cohort (n=200 patients) and validated on a testing cohort (n=25 patients), to extract 1024 deep radiomics features (d-RFs) per model. Following RNAseq of 110 samples, gene expression levels were calculated. Unsupervised classifications based on h-RFs, CAE, HSCAE and RNAseq were built with hierarchical clustering and explained according to histological features, radiological features, gene expression, pathway and survival analyses.
Results or Findings: 225 patients were included (120 men [53.3%], median age: 62 years). Three radiomics classifications were obtained (h-RF, CAE and HSCAE groups), which were not associated with the transcriptomics groups, but with prognostic radiological features known to correlate with higher grade (all P-values<0.001), and Sarculator groups (all P-values<0.001). HSCAE and h-RF groups were also associated with MFS in multivariable Cox regressions (P =0.0146 and 0.0043, respectively). Combining these groups improved the prognostic performances of the transcriptomics groups alone (c-index=0.603, increasing to 0.666 with h-RF [P=0.0380] and 0.709 with HSCAE [P=0.0110]). Fifteen genes were dysregulated and two pathways were up-regulated in the h-RF groups, which were linked to tumorigenesis and immune response.
Conclusion: Radiophenotypes of STS on pre-treatment MRI obtained with handcrafted and deep radiomics were explainable by radiologists, independently associated with MFS and strengthened transcriptomics signature.
Limitations: This is a retrospective single-centre study.
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: The study was approved by the Institutional Review Board of Bergonié Institute, comprehensive cancer centre of Bordeaux, France.

Author Block: Y. Du¹, M. Greuter², M. Prokop³, G. De Bock²; ¹Hangzhou/CN, ²Groningen/NL, ³Nijmegen/NL
Purpose: This study aimed to explore appropriate pricing for commercial deep learning computer aided detection (DL-CAD) systems in different modes of use to maximise cost savings and identify the most cost-effective mode for lung cancer screening.
Methods or Background: We evaluated DL-CAD as a concurrent, prescreening, and second reader in three representative countries. A scoping review was conducted to estimate the radiologist reading time with and without DL-CAD. The hourly cost of radiologist time was collected for the US, UK and Poland, and the monetary equivalent of saved time was calculated. The minimum number of screening CTs needed to reach break-even for a one-time investment for a DL-CAD was calculated.
Results or Findings: The mean reading time per case without DL-CAD was 2.5 minutes. It decreased by more than one minute when using DL-CAD as a concurrent and prescreening reader, respectively. It increased by about a half minute for DL-CAD as second reader. These reading times translated into costs of one to four euros per case for concurrent reading and one to six euro for prescreening reading. To reach break-even with a one-time investment for a DL-CAD, the minimum number of CT scans was about 12,000-54,000 for concurrent reader and 9,000-65,000 for prescreening reader.
Conclusion: Based on the current pricing, it is necessary for the per case cost to be significantly below €6 or for DL-CAD to be used in a high-workload setting to reach break-even in lung cancer screening. The use of DL-CAD as a prescreening reader has the greatest potential for cost savings.
Limitations: This study focused on the costs associated with DL-CAD, as it was beyond the study's scope to consider downstream costs related to diagnosis and treatment.
Funding for this study: This work is a part of NELCIN-B3 project. The NELCIN-B3 project is funded by The Royal Netherlands Academy of Arts and Sciences (Grant No. PSA_SA_BD_01) and Ministry of Science and Technology of the People's Republic of China, National Key R&D Program of China (Grant No. 2016YFE0103000).
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information: Because this article does not contain any studies on human or animal subjects, ethics committee approval was not sought.

Author Block: Z. Liu, Y. Sun, T. Tong; Shanghai/CN
Purpose: The purpose of this study was to develop and validate a fully automated deep learning model based on multimodal MRI for DFS prediction of locally advanced rectal cancer (LARC) patients treated with neoadjuvant radiotherapy and chemotherapy (nCRT).
Methods or Background: A total of 462 rectal cancer patients treated with nCRT from three centres were retrospectively enrolled, including clinical information, baseline multimodal MRI images (T2, ADC, Dapp, Kapp), and DFS data. The data from centres I and II were combined (373 cases) and randomly assigned to training, validation, and internal testing sets. The data from centre III were used as the external testing (89 cases). We developed a multitask joint survival model that simultaneously performed segmentation, risk classification, and survival prediction. These multitasks collaborated with each other, contributing to fully exploiting the key features of both imaging data and clinical data. Considering the multimodal input, an attention mechanism was introduced to efficiently capture useful information within and between modalities to minimise the impact of noise on experimental results. In addition, we experimented with conventional singletask models and compared the performance with our model.
Results or Findings: For risk classification, our model achieved significantly better performance than the singletask model, with an AUC of 0.960 vs 0.782 in the internal testing set, and remained at an AUC of 0.767 in the external testing set; for segmentation and survival prediction tasks, our model achieved a dice of 0.748 and a C-index of 0.731 in internal testing set, respectively.
Conclusion: Multitask deep learning model are expected to provide fully automated one-step prediction, contributing to optimising personalised treatment for LARC patients.
Limitations: Sample size were relatively small. Predictions were only made for DFS and not implemented for OS.
Funding for this study: Funding for this study was provided by the National Natural Science Foundation of China (81971687,82001776).
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: No information provided by the submitter.

Author Block: Z. Liu, Y. Sun, T. Tong; Shanghai/CN
Purpose: The consensus molecular subtype (CMS) is a novel classification system that reflects the genetic characteristics of the tumour. Among the four subtypes, CMS4 is associated with the worst prognosis. This study aimed to investigate whether a radiomics-based machine learning approach could predict CMS4 status in CRC patients.
Methods or Background: A total of 228 CRC cases from three centres were retrospectively included. Cases from centre I were divided into training (138 cases) and validation sets (33 cases) in an 8:2 ratio; cases from centre II and III were combined as the external testing set (57 cases). Sequencing data and baseline MRI images, including T2-weighted (T2WI) and contrast-enhanced (CE) sequences, were available for each case. The sequencing data was input into the CMS classification system to generate CMS subtype outcomes. Radiomics features from the two sets were extracted with the same parameter settings. Several machine learning algorithms were applied in sample balance, feature normalisation, feature filters, and classifier construction to explore the best-performing and most robust model for CMS4 prediction. The rad-score for each patient was calculated by the T2WI and CE models separately. The combined model was established by applying logistic regression on the results of the above two models.
Results or Findings: We found that the CE model achieved better performance than the T2 model in both the test set (0.815 vs 0.790) and external validation set (0.741 vs 0.702). After merging the two models, the predictive performance of the Merged model was further improved, with the AUCs of 0.855 and 0.759 in the test set and external validation set.
Conclusion: Multiparametric radiomics-based machine learning shows promising potential in distinguishing CMS4 from other subtypes of CRC.
Limitations: The study's relatively small sample size as well as the manual delineation of the lesions' contours were identified as limitations.
Funding for this study: Funding for this study was provided by the National Natural Science Foundation of China (81971687,82001776).
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: No information provided by the submitter.

Author Block: Z. Elkarghali, M. Hattink, O. Maxouri, S. Rostami, D. I. Rodríguez Sánchez , S. Trebeschi, R. G. H. Beets-Tan; Amsterdam/NL
Purpose: Interlesional genetic heterogeneity is an established fact in tumour biology; however, performing a biopsy on every lesion is not feasible. For the first time in the literature, we explore the use of medical image analysis techniques to quantify morphological heterogeneity between lesions as a proxy for biological heterogeneity and analyse its value in stratifying patients based on prognosis.
Methods or Background: We collected a diverse pancancer cohort of 1692 CE-CT scans with genetically proven diagnoses and performed complete 3D tumour segmentation. From each delineation (n=11,268 lesions), we derived radiomic features that fully characterise each lesion's morphology. We utilised seven similarity distance metrics (Euclidean, Chebyshev, City-Block, Minkowski, Correlation, Bray-Curtis, and Cosine) to measure the median morphological dissimilarity of lesions in a patient. Survival analysis (log-rank test) was performed to compare patients with high or low interlesional morphological heterogeneity (relative to the data set level median distance).
Results or Findings: We computed seven distance metrics for every combination of lesions within a patient and calculated the median as a patient level metric of morphological interlesional heterogeneity. Chebshev (Χ2=12.49, P =0.000408), City-Block (Χ2=12.08, P =0.000508), Euclidean (Χ2=11.64, P =0.000646), and Minkowski (Χ2=11.64, P=0.000646) distance measures could all strongly stratify patients into high- and low-risk groups. Cosine (Χ2=4.13, P=0.042) and correlation (Χ2=3.99, P=0.046) similarity metrics were also predictive to a lesser extent. Bray-Curtis distance measures could not significantly stratify patients (Χ2=3.35, P=0.067).
Conclusion: Interlesional morphological heterogeneity, as measured by radiomics and similarity distance metrics, strongly predicted overall survival.
Limitations: External validation has yet to be performed as a proof-of-concept study despite including over 1500 real-world cases.
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: IRB approval was granted: IRBd19-147.

Author Block: W. Peng, L. Wan, X. Tong, F. Yang, S. Wang, L. Li, H. Zhang; Beijing/CN
Purpose: The purpose of this study was to conduct a clinical assessment of deep learning reconstruction (DLR)-based rectal MRI in comparison to standard MRI.
Methods or Background: Patients with biopsy-proven rectal adenocarcinoma from November/2022 to May/2023 were prospectively enrolled in the study to undertake rectal MRI, including protocols using standard fast spin-echo (FSEstandard) and DLR-based accelerated FSE (FSEDL). Imaging quality including signal-noise ratio (SNR), contrast-noise ratio (CNR), as well as subjective scoring based on Likert scale were assessed by two radiologists. Diagnostic performance including T-staging, N-staging, EMVI, and MRF was further evaluated by five radiologists. The time consumed in the application of each diagnostic metric was documented for reading efficiency analysis.
Results or Findings: In total, 117 patients (77 males; age range 21 – 77 years) were enrolled in the study; 60 patients undertook radical surgery. DLR enabled a reduction of 65% in acquisition time. Moderate to excellent intra- and interreader agreement was achieved for all assessment metrics. FSEDL exhibited higher SNR, CNR, and subjective scores in noise, tumour margin clarity, visualisation of bowel wall layering and rectal mesorectal fascia, overall image quality, and diagnostic confidence (P < 0.05). FSEDL was rated higher T-staging accuracy by junior readers (reader 1, 58% vs 70%, P = 0.016; reader 3, 60% vs 76%, P = 0.021), with comparable performance in evaluating N-staging, EMVI, and MRF. No difference was found concerning diagnostic performance by senior readers (P > 0.05). FSEDL exhibited shorter diagnostic time in T-staging and overall evaluation by all readers, as well as in EMVI and MRF by junior readers (P < 0.05).
Conclusion: FSEDL is clinically feasible for rectal MRI, which could facilitate improved image quality and reading efficiency than FSE standard, while reducing 65% acquisition time. Moreover, it has potential in helping junior radiologists improve the accuracy of T-staging.
Limitations: This was a single-centre study.
Funding for this study: Funding was received from the CAMS Innovation Fund for Medical Sciences (CIFMS) [grant number 2021-I2M-C&T-A-017], Capital's Funds for Health Improvement and Research (CFH) [grant number 2022-2-4024], the National Natural Science Foundation of China [grant number 81971589], 2020 SKY Imaging Research Fund [grant number Z-2014-07-2003-01], and CAMS Innovation Fund for Medical Sciences (CIFMS) [grant number 2022-I2M-C&T-B-077].
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: This prospective study received approval from our institutional review board, and written informed consent was obtained from all participants.

Author Block: P. A. Bonaffini¹, A. Corsi², R. Muglia¹, G. Perugini³, M. Roscigno³, L. F. Da Pozzo³, P. Marra¹, S. Sironi¹; ¹Monza/IT, ²Liege/BE, ³Bergamo/IT
Purpose: The purpose of this study was to identify radiomic features potentially supporting the detection of clinically significant prostate cancer (csPC) in PI-RADS 4/5 lesions detected on 3T multiparametric MRI (mpMRI) studies.
Methods or Background: We retrospectively enrolled patients who underwent a 3T mpMRI (June 2016-March 2021) and with at least one PI-RADS 4-5 lesion (PI-RADS v2.1). Final pathological findings from fusion MRI-targeted biopsies served as ground truth. Clinical (age, PSA, PSA density) and MRI conventional parameters (prostate volume, mean ADC in circular 2D ROI) were collected. Included lesions were manually contoured on itk-SNAP on ADC maps, and axial T2 images; volumes of interest were also obtained. Radiomic features were extracted using Pyradiomics. Clinical and radiomic features best correlating with final histological results were selected. All models were assessed through 100 repetitions using 5-fold cross-validation. Sensitivity and specificity were assessed on validation samples.
Results or Findings: Among 945 patients who had undergone prostate mpMRI within the study period, 99 patients (median age 69 years) with 111 PI-RADS 4-5 lesions met the inclusion criteria. At the end of the histopathological analysis, 79 lesions (71%) were found to be csPC (GS≥7). The best predicting clinical (PSA density) and radiomic (ADC-wavelet-LLL_glrlm_LongRunHighGrayLevelEmphasis/texture feature, T2-wavelett-HHH_glszm_GrayLevelVariance and T2-wavelet-LLL_glszm_GrayLevelVariance/heterogenicity feature) multivariate model for PI-RADS 4-5 lesions obtained 79% sensitivity, 80% specificity, 91% PPV, 63% NPV and 79% accuracy. When considering only peripheral zone lesions, a multivariate model with only the same radiomic features gained 86% sensitivity, 80% specificity, 93% PPV, 70% NPV and 84% accuracy.
Conclusion: Texture and heterogeneity features extracted from 3T mpMRI T2 and ADC sequences may improve the detection of csPC in PI-RAD 4 and 5 lesions, demonstrating a better performance when considering only peripheral zone lesions and correlating also with PSA density in the zone-ignorant model.
Limitations: This was a monocentric retrospective study. The sample size was limited.
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 conducted following the Declaration of Helsinki. Patients' anonymity was granted.

Author Block: T. P. T. Hoang¹, P. Schindler², N. Börner¹, M. Masthoff², M. Seidensticker¹, J. Ricke¹, M. Ingrisch¹, O. Öcal¹, M. Wildgruber¹; ¹Munich/DE, ²Münster/DE
Purpose: The purpose of this study was to investigate the prognostic value of computed tomography (CT) derived imaging biomarkers in hepatocellular carcinoma (HCC) recurrence after liver transplantation (LT) and develop a predictive nomogram model.
Methods or Background: This retrospective study included 178 patients with histopathologically confirmed HCC who underwent liver transplantation between 2007 and 2021 at the two academic liver centres. We evaluated dedicated imaging features from baseline multiphase contrast-enhanced CT supplemented by several clinical findings and laboratory parameters. Time-to-recurrence (TTR) was estimated by Kaplan–Meier analysis. Univariable Cox proportional hazard regression and multivariable least absolute shrinkage and selection operator (LASSO) regression were used to identify independent prognostic factors for recurrence. A nomogram model was then built based on the independent factors selected through LASSO regression, to predict the probabilities of HCC recurrence at one, three, and five years.
Results or Findings: The rate of HCC recurrence after LT was 17.4% (31 of 178). The LASSO analysis revealed six independent predictors associated with an elevated risk of tumour recurrence. These predictors included the presence of peritumoural enhancement, the presence of over three tumour lesions, the largest tumour diameter exceeding 3 cm, serum alpha-fetoprotein (AFP) levels surpassing 400 ng/mL, and the presence of a tumour capsule. Conversely, a history of bridging therapies was found to be correlated with a reduced risk of HCC recurrence. In addition, Kaplan-Meier analysis with log-rank test showed patients with irregular margins, satellite nodules, or small lesions displayed significantly shorter time-to-recurrence. Our nomogram demonstrated good performance, yielding a C-index of 0.835 and AUC values of 0.86, 0.88, and 0.85 for the predictions of 1-year, 3-year, and 5-year TTR, respectively.
Conclusion: Imaging parameters derived from baseline contrast-enhanced CT showing malignant characteristics and aggressive growth patterns, along with serum AFP and a history of bridging therapies, can serve as biomarkers for predicting HCC recurrence after transplantation.
Limitations: First, this study has a limited sample size. Second, patients with various types of bridging therapies have been included. Although this situation causes inhomogeneity, it reflects the daily clinical routine of large transplantation centres.
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: This study was approved by an ethics committee; the notification can be found under the number 22-0110.

Author Block: A. Ammirabile¹, L. Viganò¹, V. Zanuso¹, F. Fiz², M. E. Laino¹, M. Francone¹, M. Sollini¹, L. Rimassa¹; ¹Milan/IT, ²Genoa/IT
Purpose: Intrahepatic cholangiocarcinoma (ICC) is an aggressive disease with increasing incidence. Comprehensive molecular profiling has shown genetic alterations that could be the target of systemic therapies. Texture analysis of imaging has led to a reliable prediction of pathology data. This study investigates whether CT-based radiomics can non-invasively predict genetic alterations in ICC.
Methods or Background: All consecutive patients eligible for systemic therapy for a mass-forming ICC (01/2016-06/2022) were considered. Inclusion criteria were: the availability of a contrast-enhanced CT at diagnosis before any treatment with an adequate quality of the portal phase for textural analyses, complete molecular profiling by NGS or FISH evaluation for FGFR2 gene fusion/rearrangement. Genetic analyses were performed on surgical specimen or biopsy. The tumour was manually segmented and radiomic features were automatically extracted using the LifeX software. Predictive models were built considering clinical and radiomic data.
Results or Findings: 90 patients were enrolled (58 NGS,32 FISH) with a median age of 65 years. The most common genetic alterations were FGFR2 (20/90), IDH1-2 (12/58), KRAS (9/58). The performances of the predictive models for FGFR2 and IDH1-2 improved by adding radiomic features to clinical data, reaching a C-index of 0.892 (vs 0.800 of the clinical model) and 0.811 (vs 0.670), respectively, at internal validation. The pure radiomic model for the prediction of KRAS mutations achieved a C-index of 0.862 at internal validation (vs 0.660 of the pure clinical model) without further improvements with the addition of clinical features.
Conclusion: The radiomic features extracted from CT at ICC diagnosis can potentially provide a reliable noninvasive prediction of its genetic status with a major impact on therapeutic strategies.
Limitations: The limitations of the study are its retrospective nature, lack of external validation as well as the commonest mutations being the subject of analysis.
Funding for this study: Funding was provided by the AIRC grant #2019−23822.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: The local review board of each centre approved the study protocol (coordinating centre approval: protocol number 142/21, date of approval 17/03/2021). Because of the retrospective nature of the study, the need for specific informed consents was waived.

Author Block: A. Ammirabile¹, F. Fiz², S. Langella³, M. Serenari⁴, M. Sollini¹, A. Chiti¹, G. Torzilli¹, F. Leva¹, L. Viganò¹; ¹Milan/IT, ²Genoa/IT, ³Turin/IT, ⁴Verona/IT
Purpose: In patients affected by intrahepatic cholangiocarcinoma (ICC), the prediction of survival based on morphological and clinical parameters has limited reliability. The present study aims to elucidate if the textural features of ICC and its peritumoural tissue extracted from preoperative computed tomography (CT) improve the prediction of survival after resection.
Methods or Background: All consecutive patients undergoing resection for ICC at six high-volume centres (2009-2019) were considered. The arterial and portal phases of CT performed <60 days before surgery were analysed. The tumour was manually segmented (tumour-VOI), a 5-mm automatic volume expansion was applied to encompass the peritumoural tissue (margin-VOI). The radiomic features were automatically extracted by the LifeX software. For overall and progression-free survival (OS/PFS), we considered pre- and post-operative predictive models, based on clinical data and radiomic features from portal and arterial phases.
Results or Findings: 215 patients were included (median age 67.5 years). The three-year OS/PFS rates were 57.0% and 34.9% (median follow-up 28 months). The predictive model of OS based on clinical variables had a C-index of 0.681. The performance progressively improved by adding the radiomic features: C-index =0.710 including portal tumor-VOI, C-index =0.752 including portal tumour-Margin-VOI; C-index=0.764 including all arterial and portal VOIs. The latter model retained clinical variables (CA19-9, tumour pattern), tumour features (density, homogeneity, GLRLM indices), and margin data (kurtosis, compacity, shape). The model had a performance equivalent to the post-operative clinical model including the pathology data (C-index =0.765). The same results were observed for PFS.
Conclusion: CT-based radiomics of ICC and peritumoural tissue improves prediction of survival, and, in combination with clinical data, leads to a preoperative estimation of outcome equivalent to the post-operative one.
Limitations: The limitations of the study are its retrospective nature, lack of external validation as well as the absence of radiomic features from CT late-phase.
Funding for this study: Funding was provided by the AIRC grant #2019−23822.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: The local review board of each centre approved the study protocol (coordinating centre approval: protocol number 142/21, date of approval 17/03/2021). Because of the retrospective nature of the study, the need for specific informed consent was waived.

Author Block: M. Polici, D. Valanzuolo, D. Pugliese, G. Tremamunno, F. Palmeri, M. Zerunian, D. De Santis, D. Caruso, A. Laghi; Rome/IT
Purpose: The study aimed to develop a radiomic model able to identify high-risk colon cancer by analysing properative CT scans.
Methods or Background: The study population included 300 patients with nonmetastatic colon cancer were retrospectively enrolled from January 2015 to June 2020. The population was divided into two groups, high-risk and no-risk, following the presence of at least one high-risk clinical factor between staging T4, LVI, PNI, budding, and nodal metastases. All patients had baseline CT scans, and 3D cancer segmentation was performed on the portal phase by two expert radiologists using open-source software (3DSlicer v4.10.2). Among the 107 radiomic features extracted, stable features were selected to evaluate the interclass correlation (ICC) (cut-off ICC > 0.8). Stable features were compared between the two groups (T-test or Mann–Whitney), and the significant features were selected for univariate and multivariate logistic regression to build a predictive radiomic model. Furthermore, survival analyses were performed with Kaplan-Meier curves, with progression within 24 months considered as end-point.
Results or Findings: In total, 210/300 were classified as high-risk and 90/300 as no-risk. A total of 27 radiomic features were stable (0.80 ≤ ICC < 0.92). Among these, 15 features were significantly different between the two groups (P < 0.05), and only eight features were selected to build the radiomic model. The radiomic model yielded an AUC of 0.73. Three radiomic features demonstrated correlation with progressive disease in Kaplan-Meier curves.
Conclusion: In conclusion, the radiomic model could be seen as a performant, non-invasive imaging tool to properly stratify colon cancers with high-risk disease especially in preoperative setting.
Limitations: The retrospective nature of the study was identified as a limitation; the study populations were unbalanced.
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: Written informed consent was acquired for all patients and Institutional Review Board approval was obtained.