Research

Efficacy Evaluation of Optimal Patient Selection for Hypopharyngeal Cancer Organ Preservation Therapy using MRI-derived Radiomic Signature: Bi-institutional Propensity Score Matched Analysis.

Shihmin Lin¹, Cheyu Hsu², Yuehchou Lee³, T. Li⁴, S. Kuo⁵, Weichung Wang⁶
Presented: October 2019, European Society for Medical Oncology (ESMO) Congress 2019

1. Radiation Oncology, Chang Gung Medical Foundation – Linkou Chang Gung Memorial Hospital, TW; 2. Department Of Oncology, National Taiwan university hostipal, TW; 3. Mathematics, National Taiwan University, TW; 4. Radiation Oncology, National Taiwan University Hospital, TW; 5. Department Of Oncology, National Taiwan University Hospital, TW; 6. Institute Of Applied Mathematical Sciences, National Taiwan University, TW Background: Early loco-regional failure (LRF) after organ preservation therapy (OPT) varies widely for hypopharyngeal squamous cell carcinoma (HPSCC) patients. We aim to develop and validate a MRI-derived radiomic signature RS for the prediction of 1year LRF in HPSCC treated with OPT, and investigate its efficacy between OPT and total laryngectomy (TL) cohort. Methods: A total of 3912 MRI-based radiomic features (RF) of pretreatment tumors were obtained from 370 HPSCC patients, including OPT cohort1 (OPT1; n = 186), OPT cohort2 (OPT2; n=88), and TL cohort (TLc; n=96). Variational autoencoder (VAE), trained with symmetric two encoded and decoded layers of neural network was applied to reduce the dimensionality of original RF to 128 VAE-RF. Least absolute shrinkage and selection operator with 10-fold cross validation performs features selection and constructs RS to predict 1-year LRF events in OPT1, which was validated in OPT2 and TLC. Harrell’s Cindex was used to evaluate the discriminative ability of RS. Optimal cut-point for dichotomized RS risk category was determined via Youden index. Pair-wise propensity score matching (caliper 0.2) using pre-treatment variables (age, gender, TNM stage) was applied to compare the impact of OPT and TL under different RS risk categories. Results: The RS yielded 1000 times bootstrapping corrected C-index of 0.753, 0.745 and 0.398 in the OPT1, OPT2 and TLC, respectively. Dichotomized risk category using Youden cut-point of RS yielded 1 year LRF predictive accuracy of 71.12%, 70.41%, and 41.74% in OPT1, OPT2 and TLC, respectively. In RS-high risk group, OPT were associated with poor progression-free survival (PFS, HR: 1.752, p=0.032), while in RS-low risk group, OPT did not deteriorate the PFS (HR: 0.774, p=0.416). Conclusions: The RS-based model provides a novel to predict 1-year LRF and survival in patients with HPSCC who received OPT. The prediction performance discrepancy of MRI-derived RS in TLC also emphasizes the role of TL in RS-high risk group.


ESMOThe European Society for Medical Oncology (ESMO) Congress is the appointment in Europe for clinicians, researchers, patient advocates, journalists and the pharmaceutical industry from all over the world to get together, learn about the latest advances in oncology and translate science into better cancer patient care. Source: ESMO Congress 2019 Website https://www.esmo.org/Conferences/Past-Conferences/ESMO-Congress-2019

Radiographic Phenotyping to Identify Intracranial Disseminated Recurrence in Brain metastases Treated With Radiosurgery Using Contrast-enhanced MR Imaging

Cheyu Hsu¹, S. Kuo², Weichung Wang³, T.W. Chen⁴, Yuehchou Lee⁵
Presented: October 2019, European Society for Medical Oncology (ESMO) Congress 2019

1. Radiation Oncology, National Taiwan university hostipal, TW; 2. Department Of Oncology, National Taiwan University Hospital, TW; 3. Institute Of Applied Mathematical Sciences, National Taiwan University, TW; 4. Department Of Oncology, College of Medicine, National Taiwan University, TW; 5. Mathematics, National Taiwan University, TW Background: Early intracranial progression (ICP) reduce the efficacy of first line radiosurgery (SRS) for brain metastases. We aim to develop and validate a MR imaging-derived radiomic signature (RS) via deep learning approach for the prediction of 1-year disseminated ICP (DICP; more than or equal to 3 lesions and leptomeningeal carcinomatosis) in brain metastases patients treated with SRS. Methods: A total of 1304 MRI-based radiomic features of pretreatment tumors were obtained from 208 patients with 451 lesions, who received first line SRS during August 2008 to January 2018. Variational autoencoder (VAE), trained with symmetric two encoded and decoded layers of neural network and 1,649,560 trainable parameters, was applied to reduce the dimensionality of radiomc features to 128 VAE-radiomic features. Penalized regression with 10-fold cross validation using least absolute shrinkage and selection operator performs features selection and construct RS to predict 1-year DICP events in train set of 150 patients, which was validated in test set of 58 patients. Harrell’s C-index was used to evaluate the discriminative ability of RS in both sets. The correlation of VAE-radiomic features and molecular features was analyzed by student t-test. Survival analysis was calculated using the Kaplan-Meier method. Results: The RS yielded 1000 times bootstrapping corrected C-index of 0.746 and 0.747 for discrimination of 1-year DICP in the train and test cohorts, respectively. As for the subgroup of patients with lung (n=175) and breast (n=23) origin, the RS also showed good predictive performance with C-indices of 0.735 and 0.755, respectively. EGFR-mutation (n=113) and ER (n=22) status were associated with selected VAE-radiomic features No. 98 (p = 0.035) and No.127 (p=0.44), respectively. Dichotomized risk category using RS of -0.769 (Youden index) as cut-off point yielded median overall survival of 57.7 months in low risk compared to 20.5 months in high risk group (p < 0.01). Conclusions: The RS model provides a novel approach to predict 1-year DICP and survival in brain metastases receiving SRS, and is warranted to be integrated into GPA for optimal selection of patients treated with first line SRS.


The European Society for Medical Oncology (ESMO) Congress is the appointment in Europe for clinicians, researchers, patient advocates, journalists and the pharmaceutical industry from all over the world to get together, learn about the latest advances in oncology and translate science into better cancer patient care. Source: ESMO Congress 2019 Website https://www.esmo.org/Conferences/Past-Conferences/ESMO-Congress-2019

Severe Stenosis Detection using 2D Convolutional Recurrent Network

Chiatse Wang¹, Chih-Kuo Lee², Yu-Cheng Huang², Wen-Jeng Lee², Tzung-Dau Wang², Weichung Wang³, Cheng-Ying Chou³, Junting Chen¹, Weidao Lee¹
Presented: September 2019, European Society of Cardiology (ESC) Congress 2019

1. Graduate Program of Data Science, National Taiwan University and Academia Sinica, TW; 2. National Taiwan University Hospital, TW; 3. National Taiwan University, TW Purpose: Stenosis detection is a critical point in the diagnosis of coronary artery disease (CAD). Manually detecting stenosis over the complete coronary artery takes about 30 minutes. Our stenoses detector can detect all stenoses that were greater than 70% in less than 20 seconds per patient. This is a critical decrease in detection time. Method: We develop a workflow to organize the raw data, perform image preprocessing including cross-sectional plane sequence generator, inference by trained models, then visualize the results. The model contains 23 thousand parameters that were constructed by a recurrent neural network (RNN) following a convolutional neural network (CNN). Our model was trained on the dataset provided by Rotterdam Coronary Artery Stenoses Detection and Quantification Evaluation Framework. To detect severe stenoses that were greater than 70%, we trained this model by classifying greater than 50% or not. This process helps to eliminate the bias caused by an imbalanced dataset. Results: The present work demonstrated the feasibility of detecting severe stenoses using deep learning-based network. Further work to elaborate on the algorithm and incorporate it into the National Taiwan University Hospital diagnostic workflow. ESC


The European Society of Cardiology (ESC) Congress is the world’s largest gathering of cardiovascular professionals contributing to global awareness of the latest clinical trials and breakthrough discoveries. ESC Congress 2019 together with World Congress of Cardiology takes place from 31 August to 4 September at the Expo Porte de Versailles in Paris, France. Source: ESC Congress Website https://www.escardio.org/Congresses-%26-Events/ESC-Congress

Constructing a Platform based on Deep Learning Model to Mimic the Self-Organization Process of CT Images Order for Automatically Recognizing Human Anatomy

Feng-Mao Lin¹, Chi-Wen Chen¹, Wei-Da Huang¹, Liangtsan Wu², Anthony Costa³, Eric K. Oermann³ and Weichung Wang*⁴
Presented: December 2019, Radiological Society of North America (RSNA) 2019

1. Cohesion Information Technology, TW; 2. Foxconn Health Technology Business Group, USA; 3. Department of Neurological Surgery, Icahn School of Medicine, USA; 4. Department of Mathematics, National Taiwan University, TW *Corresponding author Purpose: To demonstrate the ability of a deep learning application to automatically identify computed tomography (CT) slice regions by major Human anatomy. This application will be deployed in National Health Insurance of Taiwan (NHI) to classify the around 458 million CT images in 2018. Materials and Methods: 954 and 4095 of CT series were selected for training and testing correspondingly. The voxel spacing must > 0.6 mm, and each series must > 40 slices. Each image was standardized to 128^2 pixels. The AlexNet and ResNet were trained with greyscale images and the 3 color images (bone, liquid and air), correspondingly. The loss function is identical to Ke Yan, and et al. in 2018 and guides slice scores increased by slice order. Linear regression was used to adjust slice score of a series in which the r-square < 0.8. The series was split into 4 parts and a new slice score was estimated from two of the best parts. Manually annotated lung boundary was used to find the cutoff for measuring sensitivity and specificity. Results: The AlexNet and ResNet was trained for 2 days. The r-square of linear regression was to measure the linearity between the slice score and its order. The amount of series with r-square < 0.8 was reduced from 4.1% to 1.7% in AlexNet and 6.8% to 2.2% in ResNet by using our error-correction approach. Fig. 1 depicted the images with similar slice scores having a similar body part. Based on the lung boundary, the score variant of the lower boundary was larger than the upper boundary. The cutoff was selected based on the highest value of specificity*sensitivity. ResNet had the best prediction performance in training data and validation data (Spec. > 0.94, Sens. > 0.9). AlexNet provided the best prediction performance in NHI validation data (Spec. > 0.91 and Sens. > 0.94). The error correction slightly improved sensitivity and specificity. The specificity and sensitivity were both larger than 0.9 in NHI validation data by using AlexNet and ResNet. Conclusion and Discussion: First, the preprocessing process could accelerate the training process and reach lower losses by using ResNet and AlexNet is efficient during the prediction. Fig 2. showed our error correction process successfully adjusting the slice score to the corresponding body part. Since the organ boundary was varied from person to person, this approach is good for large part Identification. Although we found ResNet and error correction could provide good prediction quality with small training data, the model proposed by Ke Yan, and et al. in 2018 trained with large training data is one of the states of art methods. Clinical relevance / applications: NHI collected around 458 million medical CT images in 2018. Our application will deploy in one of the largest medical databases in the world. Precisely retrieve the certain images of Human Anatomy could accelerate related application development and reduce storage usage. Acknowledgment: We appreciate that the NHI Artificial Intelligence Application Service Trial provided valuable data for our model validation in Taiwan. RSNA


The Radiological Society of North America (RSNA) is a non-profit organization with over 54,000 members from 136 countries around the world. It provides high-quality educational resources, including continuing education credits toward physicians’ certification maintenance, host the world’s largest radiology conference and publish two top peer-reviewed journals: Radiology and RadioGraphics. Source: RSNA 2019 Website https://rsna2019.rsna.org/

Differentiation between Pancreatic Cancer and Nontumorous Pancreas on Computed Tomography by Radiomics and Machine Learning

Po-Ting Chen¹, Huihsuan Yen², Dawei Chang³, Wei-Chih Liao⁴, Kao-Lang Liu¹, Holger R. Roth⁵, Weichung Wang³, Tinghui Wu³
Presented: December 2019, Radiological Society of North America (RSNA) 2019

1. Medical Imaging, National Taiwan University Hospital and National Taiwan University College of Medicine, TW; 2. Data Science Degree Program, National Taiwan University & Academia Sinica, TW; 3. Institute of Applied Mathematical Sciences, National Taiwan University, TW; 4. Department of Internal Medicine, National Taiwan University Hospital and National Taiwan University College of Medicine, TW 5. NVIDIA, USA Purpose: Pancreatic cancer (PC) is the most lethal cancer and the fourth leading cause of cancer deaths in the United States. Radiomics is a methodology that extracts quantitative statistics and features from medical images. The purpose of this study is to develop a machine learning model to differentiate PC from nontumorous pancreas (NP) on contrast-enhanced computed tomography (CT) using radiomic features. Materials and Methods: Contrast-enhanced venous phase CT images of 100 cases with PC and 100 controls were reviewed by an expert radiologist. The tumors and pancreas in PC cases were manually labeled by the radiologist, whereas the pancreas was segmented by a pre-trained deep learning model in some of the NP cases. Data were split into training set (60 NP cases, 60 PC cases), validation set (20 NP cases, 19 PC cases), and test set (20 NP cases, 19 PC cases). Pancreas and tumors were cut into patches of 20 pixels by 20 pixels for subsequent extraction of radomic features. A total of 91 radiomic features were extracted and subject to eXtreme Gradient Boosting (XGBoost) model to perform classification. Results: A total of 3596 patches of PC and 19446 patches of NP were generated and used for training, and the testing set included 691 patches of PC and 3889 patches of NP. For differentiation between PC and NP, the accuracy of the XGBoost by patch-based analysis was 93.43%, with an area under the receiver operating characteristic (ROC) curve (AUC) of 0.94712. In patient-based analysis, the accuracy, sensitivity, specificity and AUC were 95.12%, 0.90476, 1, and 0.95238, respectively. Top 10 features with highest feature importance score were median, 10 percentile, energy, skewness, 90 percentile, maximum, minimum, and kurtosis in first order statistics, dependence nonuniformity in gray level dependence matrix (GLDM), and cluster shade in gray level cooccurrence matrix (GLCM). Conclusion: We developed a machine learning model that could differentiate between CTs of pancreas with PC and without PC with a 95.12% accuracy in patient-based analysis and 93.43% accuracy in patch-based analysis. Among the important features which our model selects, features in first order statistics have the highest importance score followed by features in higher order statistics related to nonuniformity. CLINICAL RELEVANCE/APPLICATION: This model can accurately differentiate between cancerous and nontumorous pancreas and is a potential computer-aided diagnosis tool. RSNA


The Radiological Society of North America (RSNA) is a non-profit organization with over 54,000 members from 136 countries around the world. It provides high-quality educational resources, including continuing education credits toward physicians’ certification maintenance, host the world’s largest radiology conference and publish two top peer-reviewed journals: Radiology and RadioGraphics. Source: RSNA 2019 Website https://rsna2019.rsna.org/

Differentiation Between Pancreatic Cancer and Normal Pancreas on Computed Tomography with Artificial Intelligence

Wei-Chih Liao¹, Wei-Chung Wang², Ting-Hui Wi², Kao-Lang Liu³, Po-Ting Chen³, Hui-Hsuan Yen⁴, Holger R. Roth⁵
Presented: May 2019, Digestive Disease Week® (DDW) 2019

1. Department of Internal Medicine, National Taiwan University Hospital and National Taiwan University College of Medicine, TW; 2. Institute of Applied Mathematical Sciences, National Taiwan University, TW; 3. Medical Imaging, National Taiwan University Hospital and National Taiwan University College of Medicine, TW; 4. Data Science Degree Program, National Taiwan University & Academia Sinica, TW; 5. NVIDIA, USA Background: Computed tomography (CT) is the major modality for detection and evaluation of pancreatic cancer (PC). However, approximately one-third of PCs <2 cm are missed by CT, and differentiation between PCs and benign pancreatic lesions are often challenging. Aim: To develop and test a deep neural network to differentiate between PC and normal pancreas. Methods: CT images of 70 patients with histologically confirmed pancreatic adenocarcinoma and 70 control subjects with normal pancreas were extracted from archive. Images of 10 patients and 10 controls were randomly selected as test set, and others were used as training set. The pancreas and tumor were manually labeled by a radiologist experienced in pancreatic imaging and served as ground truth. Each image was preprocessed by windowing and normalization before cutting into patches of 50 pixels in length and width. Patches in which cancer occupied >30% of the area were labeled as cancer, whereas patches that did not contain cancer and included pancreas occupying>50% of the area, either from controls or PC patients, were labeled as normal pancreas. A convolution neural network (CNN) was trained using one TITAN V (NVIDIA) to classify patches into cancer or normal using binary cross-entropy as loss function, and its performance was assessed on the test set. For patient-based analysis, normal patches from PC patients were excluded, and patients were classified as having PC if more than 50% of his/her patches were predicted as cancer by CNN. Results: A total of 2522 patches of PC and 3808 patches of normal pancreas were generated and used for training, and the test set included 533 patches of PC and 949 patches of normal pancreas. For differentiation between PC and normal pancreas, the accuracy of the CNN by patch-based analysis was 77.1%, with an area under the receiver operating characteristic (ROC) curve (AUC) of 0.80 (Table and Figure). In patient-based analysis, the accuracy and AUC were 90% and 0.96, respectively. The computation time for analyzing the whole test set (1482 patches) was less than 5 seconds. Conclusion: We developed a CNN that could differentiate between CTs of pancreas with PC and without PC with a 90% accuracy in patient-based analysis and 77.1% accuracy in patch-based analysis. This deep-learning model is a potential computer-aided diagnosis tool to facilitate early and accurate diagnosis of PC. DDW Figure legend: Differentiation between normal pancreas and pancreatic cancer by CNN. A. Patches from CT of normal pancreas and pancreatic cancer. B. ROC curve of patch-based analysis in test set. C. ROC curve of patient-based analysis in test set. DDW


Digestive Disease Week® (DDW) is the world’s largest gathering of physicians, researchers and industry in the fields of gastroenterology, hepatology, endoscopy and gastrointestinal surgery and is recognized as one of the top 50 medical meetings by HCEA. Source: DDW Homepage https://ddw.org/

Deep learning to distinguish pancreatic cancer tissue from non-cancerous pancreatic tissue: a retrospective study with cross-racial external validation.

Kao-Lang Liu, Tinghui Wu, Po-Ting Chen, Yuhsiang M. Tsai, Holger Roth, Ming-Shiang Wu, Wei-Chih Liao, and Weichung Wang
Presented: June 2020, The Lancet Digital Health

Background

The diagnostic performance of CT for pancreatic cancer is interpreter-dependent, and approximately 40% of tumours smaller than 2 cm evade detection. Convolutional neural networks (CNNs) have shown promise in image analysis, but the networks’ potential for pancreatic cancer detection and diagnosis is unclear. We aimed to investigate whether CNN could distinguish individuals with and without pancreatic cancer on CT, compared with radiologist interpretation.

Methods

In this retrospective, diagnostic study, contrast-enhanced CT images of 370 patients with pancreatic cancer and 320 controls from a Taiwanese centre were manually labelled and randomly divided for training and validation (295 patients with pancreatic cancer and 256 controls) and testing (75 patients with pancreatic cancer and 64 controls; local test set 1). Images were preprocessed into patches, and a CNN was trained to classify patches as cancerous or non-cancerous. Individuals were classified as with or without pancreatic cancer on the basis of the proportion of patches diagnosed as cancerous by the CNN, using a cutoff determined using the training and validation set. The CNN was further tested with another local test set (101 patients with pancreatic cancers and 88 controls; local test set 2) and a US dataset (281 pancreatic cancers and 82 controls). Radiologist reports of pancreatic cancer images in the local test sets were retrieved for comparison.

Findings

Between Jan 1, 2006, and Dec 31, 2018, we obtained CT images. In local test set 1, CNN-based analysis had a sensitivity of 0·973, specificity of 1·000, and accuracy of 0·986 (area under the curve [AUC] 0·997 (95% CI 0·992–1·000). In local test set 2, CNN-based analysis had a sensitivity of 0·990, specificity of 0·989, and accuracy of 0·989 (AUC 0·999 [0·998–1·000]). In the US test set, CNN-based analysis had a sensitivity of 0·790, specificity of 0·976, and accuracy of 0·832 (AUC 0·920 [0·891–0·948)]. CNN-based analysis achieved higher sensitivity than radiologists did (0·983 vs 0·929, difference 0·054 [95% CI 0·011–0·098]; p=0·014) in the two local test sets combined. CNN missed three (1·7%) of 176 pancreatic cancers (1·1–1·2 cm). Radiologists missed 12 (7%) of 168 pancreatic cancers (1·0–3·3 cm), of which 11 (92%) were correctly classified using CNN. The sensitivity of CNN for tumours smaller than 2 cm was 92·1% in the local test sets and 63·1% in the US test set.

Interpretation

CNN could accurately distinguish pancreatic cancer on CT, with acceptable generalisability to images of patients from various races and ethnicities. CNN could supplement radiologist interpretation.

Funding

Taiwan Ministry of Science and Technology