Deep learning models for radiography body-part classification and chest radiograph projection/orientation classification: a multi-institutional study.

To develop and validate a deep learning (DL) model based on quantitative analysis of contrast-enhanced ultrasound (CEUS) images that predicts early recurrence (ER) after thermal ablation (TA) of colorectal cancer liver metastasis (CRLM). Between January 2010 and May 2019, a total of 207 consecutive patients with CRLM with 13,248 slice images at three dynamic phases who received CEUS within 2 weeks before TA were retrospectively enrolled in two centres (153 for the training cohort (TC), 32 for the internal test cohort (ITC), and 22 for the external test cohort (ETC)). Clinical and CEUS data were used to develop and validate the clinical model, DL model, and DL combining with clinical (DL-C) model to predict ER after TA. The performance of these models was compared by the receiver operating characteristic curve (ROC) with the DeLong test. After a median follow-up of 56 months, 49% (99/207) of patients experienced ER. Three key clinical features (preoperative chemotherapy (PC), lymph node metastasis of the primary colorectal cancer (LMPCC), and T stage) were used to develop the clinical model. The DL model yielded better performance than the clinical model in the ETC (AUC: 0.67 for the clinical model, 0.76 for the DL model). The DL-C model significantly outperformed the clinical model and DL model (AUC: 0.78 for the DL-C model in the ETC; both, p < 0.001). The model based on CEUS can achieve satisfactory prediction and assist physicians during the therapeutic decision-making process in clinical practice. • This is an exploratory study in which ablation-related contrast-enhanced ultrasound (CEUS) data from consecutive patients with colorectal cancer liver metastasis (CRLM) were collected simultaneously at multiple institutions. • The deep learning combining with clinical (DL-C) model provided desirable performance for the prediction of early recurrence (ER) after thermal ablation (TA). • The DL-C model based on CEUS provides guidance for TA indication selection and making therapeutic decisions.

e13628 Background: Despite the wide availability and cost-effectiveness of chest X-rays (CXRs), their potential as a biomarker for immune checkpoint inhibitor (ICI) response in non-small cell lung cancer (NSCLC) remains largely unexplored. We developed and validated a novel deep learning model that predicts ICI response from a single pre-treatment CXR, independent of histopathologic biomarkers including PD-L1 expression and immune phenotype. Methods: A retrospective dataset of 2,905 NSCLC patients at Samsung Medical Center was collected, including pretreatment CXRs and H&amp;E whole slide images (WSI). A deep learning model was trained using 1,939 pre-ICI CXRs to predict ICI outcomes, with time to next treatment (TTNT) within 6 months as a proxy. To evaluate regimen-specific performance, the model was tested on pretreatment CXRs from three patient groups defined by first-line regimen: A (N = 119, ICI monotherapy), B (N = 327, platinum-based chemotherapy), and C (N = 520, ICI-chemotherapy combination). To interpret the model's behavior, we generated segmentation masks for tumor regions and lung fields using a separate model and injected white noise at various signal-to-noise ratios (SNR). We assessed tumor infiltrating lymphocyte (TIL) density and immune phenotype on 113 paired H&amp;E WSIs from test set A using Lunit SCOPE IO, a deep learning H&amp;E analyzer, to correlate CXR predictions with the tumor microenvironment. Results: The model achieved an area under the ROC curve (AUC) of 0.745 in test set A, with predicted good responders showing significantly prolonged TTNT (median: 12.0 vs 2.0 months, hazard ratio [HR] 0.41, p &lt; 0.001) and OS (median: 24.7 vs 6.0 months, HR 0.42, p &lt; 0.001) compared to predicted non-responders. The model's AUC was 0.531 in test set B (predicting TTNT after chemotherapy) and 0.654 in test set C (ICI-chemotherapy combination), demonstrating specificity of the model to ICI. Noise injection in the nodule region and lung field led to a decrease in AUC, with the most significant decrease in extra-tumor regions of the ipsilateral lung. Prediction scores were not correlated with intratumoral TIL density (r = 0.07, p = 0.44) or PD-L1 expression (r = -0.08, p = 0.56). The prediction of the CXR model for ICI monotherapy (test set A) was independent of high PD-L1 expression (PD-L1≥50%, HR 0.43, p = 0.02; PD-L1 &lt; 50%, HR 0.38, p = 0.13) or immune phenotypes determined by TIL enrichment (Inflamed: HR 0.48, p &lt; 0.05; Non-inflamed: HR 0.38, p &lt; 0.001). Conclusions: A deep learning CXR model predicted ICI response in NSCLC regardless of PD-L1 expression and immune phenotype. This work suggests that CXRs contain intrinsic information relevant to ICI response, achieving competitive performance with state-of-the-art CT or histology based approaches despite relying on broader lung features.

COVID-19 is highly infectious and causes acute respiratory disease. Machine learning (ML) and deep learning (DL) models are vital in detecting disease from computerized chest tomography (CT) scans. The DL models outperformed the ML models. For COVID-19 detection from CT scan images, DL models are used as end-to-end models. Thus, the performance of the model is evaluated for the quality of the extracted feature and classification accuracy. There are four contributions included in this work. First, this research is motivated by studying the quality of the extracted feature from the DL by feeding these extracted to an ML model. In other words, we proposed comparing the end-to-end DL model performance against the approach of using DL for feature extraction and ML for the classification of COVID-19 CT scan images. Second, we proposed studying the effect of fusing extracted features from image descriptors, e.g., Scale-Invariant Feature Transform (SIFT), with extracted features from DL models. Third, we proposed a new Convolutional Neural Network (CNN) to be trained from scratch and then compared to the deep transfer learning on the same classification problem. Finally, we studied the performance gap between classic ML models against ensemble learning models. The proposed framework is evaluated using a CT dataset, where the obtained results are evaluated using five different metrics The obtained results revealed that using the proposed CNN model is better than using the well-known DL model for the purpose of feature extraction. Moreover, using a DL model for feature extraction and an ML model for the classification task achieved better results in comparison to using an end-to-end DL model for detecting COVID-19 CT scan images. Of note, the accuracy rate of the former method improved by using ensemble learning models instead of the classic ML models. The proposed method achieved the best accuracy rate of 99.39%.

PurposeTo conduct a head-to-head comparison between deep learning (DL) and radiomics models across institutions for predicting microvascular invasion (MVI) in hepatocellular carcinoma (HCC) and to investigate the model robustness and generalizability through rigorous internal and external validation.MethodsThis retrospective study included 2304 preoperative images of 576 HCC lesions from two centers, with MVI status determined by postoperative histopathology. We developed DL and radiomics models for predicting the presence of MVI using B-mode ultrasound, contrast-enhanced ultrasound (CEUS) at the arterial, portal, and delayed phases, and a combined modality (B + CEUS). For radiomics, we constructed models with enlarged vs. original regions of interest (ROIs). A cross-validation approach was performed by training models on one center’s dataset and validating the other, and vice versa. This allowed assessment of the validity of different ultrasound modalities and the cross-center robustness of the models. The optimal model combined with alpha-fetoprotein (AFP) was also validated. The head-to-head comparison was based on the area under the receiver operating characteristic curve (AUC).ResultsThirteen DL models and 25 radiomics models using different ultrasound modalities were constructed and compared. B + CEUS was the optimal modality for both DL and radiomics models. The DL model achieved AUCs of 0.802–0.818 internally and 0.667–0.688 externally across the two centers, whereas radiomics achieved AUCs of 0.749–0.869 internally and 0.646–0.697 externally. The radiomics models showed overall improvement with enlarged ROIs (P < 0.05 for both CEUS and B + CEUS modalities). The DL models showed good cross-institutional robustness (P > 0.05 for all modalities, 1.6–2.1% differences in AUC for the optimal modality), whereas the radiomics models had relatively limited robustness across the two centers (12% drop-off in AUC for the optimal modality). Adding AFP improved the DL models (P < 0.05 externally) and well maintained the robustness, but did not benefit the radiomics model (P > 0.05).ConclusionCross-institutional validation indicated that DL demonstrated better robustness than radiomics for preoperative MVI prediction in patients with HCC, representing a promising solution to non-standardized ultrasound examination procedures.

Electricity load forecasting is a crucial undertaking within all the deregulated markets globally. Among the research challenges on a global scale, the investigation of deep transfer learning (DTL) in the field of electricity load forecasting represents a fundamental effort that can inform artificial intelligence applications in general. In this paper, a comprehensive study is reported regarding day-ahead electricity load forecasting. For this purpose, three sequence-to-sequence (Seq2seq) deep learning (DL) models are used, namely the multilayer perceptron (MLP), the convolutional neural network (CNN) and the ensemble learning model (ELM), which consists of the weighted combination of the outputs of MLP and CNN models. Also, the study focuses on the development of different forecasting strategies based on DTL, emphasizing the way the datasets are trained and fine-tuned for higher forecasting accuracy. In order to implement the forecasting strategies using deep learning models, load datasets from three Greek islands, Rhodes, Lesvos, and Chios, are used. The main purpose is to apply DTL for day-ahead predictions (1–24 h) for each month of the year for the Chios dataset after training and fine-tuning the models using the datasets of the three islands in various combinations. Four DTL strategies are illustrated. In the first strategy (DTL Case 1), each of the three DL models is trained using only the Lesvos dataset, while fine-tuning is performed on the dataset of Chios island, in order to create day-ahead predictions for the Chios load. In the second strategy (DTL Case 2), data from both Lesvos and Rhodes concurrently are used for the DL model training period, and fine-tuning is performed on the data from Chios. The third DTL strategy (DTL Case 3) involves the training of the DL models using the Lesvos dataset, and the testing period is performed directly on the Chios dataset without fine-tuning. The fourth strategy is a multi-task deep learning (MTDL) approach, which has been extensively studied in recent years. In MTDL, the three DL models are trained simultaneously on all three datasets and the final predictions are made on the unknown part of the dataset of Chios. The results obtained demonstrate that DTL can be applied with high efficiency for day-ahead load forecasting. Specifically, DTL Case 1 and 2 outperformed MTDL in terms of load prediction accuracy. Regarding the DL models, all three exhibit very high prediction accuracy, especially in the two cases with fine-tuning. The ELM excels compared to the single models. More specifically, for conducting day-ahead predictions, it is concluded that the MLP model presents the best monthly forecasts with MAPE values of 6.24% and 6.01% for the first two cases, the CNN model presents the best monthly forecasts with MAPE values of 5.57% and 5.60%, respectively, and the ELM model achieves the best monthly forecasts with MAPE values of 5.29% and 5.31%, respectively, indicating the very high accuracy it can achieve.

Deep learning techniques have recently demonstrated remarkable success in numerous domains. Typically, the success of these deep learning models is measured in terms of performance metrics such as accuracy and mean average precision (mAP). Generally, a model's high performance is highly valued, but it frequently comes at the expense of substantial energy costs and carbon footprint emissions during the model building step. Massive emission of CO2 has a deleterious impact on life on earth in general and is a serious ethical concern that is largely ignored in deep learning research. In this article, we mainly focus on environmental costs and the means of mitigating carbon footprints in deep learning models, with a particular focus on models created using knowledge distillation (KD). Deep learning models typically contain a large number of parameters, resulting in a 'heavy' model. A heavy model scores high on performance metrics but is incompatible with mobile and edge computing devices. Model compression techniques such as knowledge distillation enable the creation of lightweight, deployable models for these low-resource devices. KD generates lighter models and typically performs with slightly less accuracy than the heavier teacher model (model accuracy by the teacher model on CIFAR 10, CIFAR 100, and TinyImageNet is 95.04%, 76.03%, and 63.39%; model accuracy by KD is 91.78%, 69.7%, and 60.49%). Although the distillation process makes models deployable on low-resource devices, they were found to consume an exorbitant amount of energy and have a substantial carbon footprint (15.8, 17.9, and 13.5 times more carbon compared to the corresponding teacher model). The enormous environmental cost is primarily attributable to the tuning of the hyperparameter, Temperature (τ). In this article, we propose measuring the environmental costs of deep learning work (in terms of GFLOPS in millions, energy consumption in kWh, and CO2 equivalent in grams). In order to create lightweight models with low environmental costs, we propose a straightforward yet effective method for selecting a hyperparameter (τ) using a stochastic approach for each training batch fed into the models. We applied knowledge distillation (including its data-free variant) to problems involving image classification and object detection. To evaluate the robustness of our method, we ran experiments on various datasets (CIFAR 10, CIFAR 100, Tiny ImageNet, and PASCAL VOC) and models (ResNet18, MobileNetV2, Wrn-40-2). Our novel approach reduces the environmental costs by a large margin by eliminating the requirement of expensive hyperparameter tuning without sacrificing performance. Empirical results on the CIFAR 10 dataset show that the stochastic technique achieves an accuracy of 91.67%, whereas tuning achieves an accuracy of 91.78%-however, the stochastic approach reduces the energy consumption and CO2 equivalent each by a factor of 19. Similar results have been obtained with CIFAR 100 and TinyImageNet dataset. This pattern is also observed in object detection classification on the PASCAL VOC dataset, where the tuning technique performs similarly to the stochastic technique, with a difference of 0.03% mAP favoring the stochastic technique while reducing the energy consumptions and CO2 emission each by a factor of 18.5.

We developed a multimodal ultrasound (US) deep learning (DL) fusion model to automatically classify early fibrosis in patients with chronic kidney disease (CKD). This prospective study included patients with CKD who underwent continuous gray-scale US, superb microvascular imaging, and strain elastography from May to November 2022. According to the pathological tubular atrophy and interstitial fibrosis score, patients were divided into minimal and mild groups (affected area ≤10% and 11 − 25% of the total cortical volume, respectively). The dataset was divided into training (70%) and test (30%) sets. A DL model combining the features of the three US modes was developed to predict early fibrosis in patients with CKD. We compared these findings with the area under the receiver operating characteristic curve (AUC) of the clinical model by analyzing the receiver operating characteristic curve in the test set. The AUC of single-mode DL based on gray-scale US, superb microvascular imaging, and strain elastography was 0.682, 0.745, and 0.648, respectively, while that of the multimodal US DL model was 0.86. The accuracy, specificity, and sensitivity of the multimodal US DL model were 0.779, 0.767, and 0.796, respectively, and the negative and positive predictive values were 0.842 and 0.706, respectively. The AUC of the multimodal US DL model was significantly better than that of the single-mode DL and clinical models. The DL algorithm developed using multimodal US images can effectively predict early fibrosis in patients with CKD with significantly greater accuracy than single-mode DL or clinical models.

We aimed at comparing intratumoral and peritumoral deep learning, radiomics, and fusion models in predicting KRAS mutations in rectal cancer using endorectal ultrasound imaging. This study included 304 patients with rectal cancer from Fujian Medical University Union Hospital. The patients were randomly divided into a training group (213 patients) and a test group (91 patients) at a 7:3 ratio. Radiomics and deep learning models were established using primary tumor and peritumoral images. In the optimally performing regions-of-interest, two fusion strategies, a feature-based and a decision-based model, were employed to build the fusion models. The Shapley additive explanation (SHAP) method was used to evaluate the significance of features in the optimal radiomics, deep learning, and fusion models. The performance of each model was assessed using the area under the receiver operating characteristic curve (AUC) and decision curve analysis (DCA). In the test cohort, both the radiomics and deep learning models exhibited optimal performance with a 10-pixel patch extension, yielding AUC values of 0.824 and 0.856, respectively. The feature-based DLRexpand10_FB model attained the highest AUC (0.896) across all study sets. In addition, the DLRexpand10_FB model demonstrated excellent sensitivity, specificity, and DCA. SHAP analysis underscored the deep learning feature (DL_1) as the most significant factor in the hybrid model. The feature-based fusion model DLRexpand10_FB can be employed to predict KRAS gene mutations based on pretreatment endorectal ultrasound images of rectal cancer. The integration of peritumoral regions enhanced the predictive performance of both the radiomics and deep learning models.

Deep learning breast cancer risk models demonstrate improved accuracy compared with traditional risk models but have not been prospectively tested. We compared the accuracy of a deep learning risk score derived from the patient's prior mammogram to traditional risk scores to prospectively identify patients with cancer in a cohort due for screening. We collected data on 119 139 bilateral screening mammograms in 57 617 consecutive patients screened at 5 facilities between September 18, 2017, and February 1, 2021. Patient demographics were retrieved from electronic medical records, cancer outcomes determined through regional tumor registry linkage, and comparisons made across risk models using Wilcoxon and Pearson χ2 2-sided tests. Deep learning, Tyrer-Cuzick, and National Cancer Institute Breast Cancer Risk Assessment Tool (NCI BCRAT) risk models were compared with respect to performance metrics and area under the receiver operating characteristic curves. Cancers detected per thousand patients screened were higher in patients at increased risk by the deep learning model (8.6, 95% confidence interval [CI] = 7.9 to 9.4) compared with Tyrer-Cuzick (4.4, 95% CI = 3.9 to 4.9) and NCI BCRAT (3.8, 95% CI = 3.3 to 4.3) models (P &lt; .001). Area under the receiver operating characteristic curves of the deep learning model (0.68, 95% CI = 0.66 to 0.70) was higher compared with Tyrer-Cuzick (0.57, 95% CI = 0.54 to 0.60) and NCI BCRAT (0.57, 95% CI = 0.54 to 0.60) models. Simulated screening of the top 50th percentile risk by the deep learning model captured statistically significantly more patients with cancer compared with Tyrer-Cuzick and NCI BCRAT models (P &lt; .001). A deep learning model to assess breast cancer risk can support feasible and effective risk-based screening and is superior to traditional models to identify patients destined to develop cancer in large screening cohorts.

The past few years have seen a surge of applying Deep Learning (DL) models for a wide array of tasks such as image classification, object detection, machine translation, etc. While DL models provide an opportunity to solve otherwise intractable tasks, their adoption relies on them being optimized to meet latency and resource requirements. Benchmarking is a key step in this process but has been hampered in part due to the lack of representative and up-to-date benchmarking suites. This is exacerbated by the fast-evolving pace of DL models. This paper proposes DLBricks, a composable benchmark generation design that reduces the effort of developing, maintaining, and running DL benchmarks on CPUs. DLBricks decomposes DL models into a set of unique runnable networks and constructs the original model's performance using the performance of the generated benchmarks. DLBricks leverages two key observations: DL layers are the performance building blocks of DL models and layers are extensively repeated within and across DL models. Since benchmarks are generated automatically and the benchmarking time is minimized, DLBricks can keep up-to-date with the latest proposed models, relieving the pressure of selecting representative DL models. Moreover, DLBricks allows users to represent proprietary models within benchmark suites. We evaluate DLBricks using $50$ MXNet models spanning $5$ DL tasks on $4$ representative CPU systems. We show that DLBricks provides an accurate performance estimate for the DL models and reduces the benchmarking time across systems (e.g. within $95\%$ accuracy and up to $4.4\times$ benchmarking time speedup on Amazon EC2 c5.xlarge).

Explainable deep learning diagnostic system for prediction of lung disease from medical images

ObjectivesIn the midst of the coronavirus disease 2019 (COVID-19) outbreak, chest X-ray (CXR) imaging is playing an important role in diagnosis and monitoring of patients with COVID-19. We propose a deep learning model for detection of COVID-19 from CXRs, as well as a tool for retrieving similar patients according to the model’s results on their CXRs. For training and evaluating our model, we collected CXRs from inpatients hospitalized in four different hospitals.MethodsIn this retrospective study, 1384 frontal CXRs, of COVID-19 confirmed patients imaged between March and August 2020, and 1024 matching CXRs of non-COVID patients imaged before the pandemic, were collected and used to build a deep learning classifier for detecting patients positive for COVID-19. The classifier consists of an ensemble of pre-trained deep neural networks (DNNS), specifically, ReNet34, ReNet50¸ ReNet152, and vgg16, and is enhanced by data augmentation and lung segmentation. We further implemented a nearest-neighbors algorithm that uses DNN-based image embeddings to retrieve the images most similar to a given image.ResultsOur model achieved accuracy of 90.3%, (95% CI: 86.3–93.7%) specificity of 90% (95% CI: 84.3–94%), and sensitivity of 90.5% (95% CI: 85–94%) on a test dataset comprising 15% (350/2326) of the original images. The AUC of the ROC curve is 0.96 (95% CI: 0.93–0.97).ConclusionWe provide deep learning models, trained and evaluated on CXRs that can assist medical efforts and reduce medical staff workload in handling COVID-19.Key Points• A machine learning model was able to detect chest X-ray (CXR) images of patients tested positive for COVID-19 with accuracy and detection rate above 90%.• A tool was created for finding existing CXR images with imaging characteristics most similar to a given CXR, according to the model’s image embeddings.Supplementary InformationThe online version contains supplementary material available at 10.1007/s00330-021-08050-1.

This study aims to develop and validate an interpretable deep learning (DL) model and a nomogram based on endoscopic ultrasound (EUS) images for the prediction of pathological grading in pancreatic neuroendocrine tumors (PNETs). This multicenter retrospective study included 108 patients with PNETs, who were divided into train (n = 81, internal center) and test cohorts (n = 27, external centers). Univariate and multivariate logistic regression were used for screening demographic characteristics and EUS semantic features. Deep transfer learning was employed using a pre-trained ResNet18 model to extract features from EUS images. Feature selection was conducted using the least absolute shrinkage and selection operator (LASSO), and various machine learning algorithms were utilized to construct DL models. The optimal model was then integrated with clinical features to develop a nomogram. The performance of the model was assessed using the area under the curve (AUC), calibration curves, decision curve analysis (DCA), and clinical impact curves (CIC). The nomogram, which integrates the optimal DL model (Naive Bayes) with clinical features, achieved AUC values of 0.928 (95% CI 0.849–0.981) in the train cohort and 0.882 (95% CI 0.778–0.954) in the test cohort. Calibration curves revealed minimal discrepancies between predicted and actual probabilities, with mean absolute errors of 4.5% and 6.6% in the train and test cohorts, respectively. DCA and CIC demonstrated substantial net benefit and clinical utility. The SHapley Additive exPlanations (SHAP) method provided insights into the contribution of each DL feature to the model’s predictions. This study developed and validated a novel interpretable DL model and nomogram using EUS images and machine learning, which holds promise for enhancing the clinical application of EUS in identifying PNETs’ pathological grading.

Research in many fields has shown that transfer learning (TL) is well-suited to improve the performance of deep learning (DL) models in datasets with small numbers of samples. This empirical success has triggered interest in the application of TL to cognitive decoding analyses with functional neuroimaging data. Here, we systematically evaluate TL for the application of DL models to the decoding of cognitive states (e.g., viewing images of faces or houses) from whole-brain functional Magnetic Resonance Imaging (fMRI) data. We first pre-train two DL architectures on a large, public fMRI dataset and subsequently evaluate their performance in an independent experimental task and a fully independent dataset. The pre-trained DL models consistently achieve higher decoding accuracies and generally require less training time and data than model variants that were not pre-trained, while also outperforming linear baseline models trained from scratch, clearly underlining the benefits of pre-training. We demonstrate that these benefits arise from the ability of the pre-trained models to reuse many of their learned features when training with new data, providing deeper insights into the mechanisms giving rise to the benefits of pre-training. Yet, we also surface nuanced challenges for whole-brain cognitive decoding with DL models when interpreting the decoding decisions of the pre-trained models, as these have learned to utilize the fMRI data in unforeseen and counterintuitive ways to identify individual cognitive states.

Recently, the COVID-19 pandemic has emerged as one of the world's most critical public health concerns. One of the biggest problems in the present COVID-19 outbreak is the difficulty of accurately separating COVID-19 cases from non-COVID-19 cases at an affordable price and in the initial stages. Besides the use of antigen Rapid Test Kit (RTK) and Reverse Transcription Polymerase Chain Reaction (RT-PCR), chest x-rays (CXR) can also be used to identify COVID-19 patients. Unfortunately, manual checks may produce inaccurate results, delay treatment or even be fatal. Because of differences in perception and experience, the manual method can be chaotic and imprecise. Technology has progressed to the point where we can solve this problem by training a Deep Learning (DL) model to distinguish the normal and COVID-19 X-rays. In this work, we choose the Convolutional Neural Network (CNN) as our DL model and train it using Kaggle datasets that include both COVID-19 and normal CXR data. The developed CNN model is then deployed on the website after going through a training and validation process. The website layout is straightforward to navigate. A CXR can be uploaded and a prediction made with minimal effort from the patient. The website assists in determining whether they have been exposed to COVID-19 or not.