RESUMO
Reliability learning and interpretable decision-making are crucial for multi-modality medical image segmentation. Although many works have attempted multi-modality medical image segmentation, they rarely explore how much reliability is provided by each modality for segmentation. Moreover, the existing approach of decision-making such as the softmax function lacks the interpretability for multi-modality fusion. In this study, we proposed a novel approach named contextual discounted evidential network (CDE-Net) for reliability learning and interpretable decision-making under multi-modality medical image segmentation. Specifically, the CDE-Net first models the semantic evidence by uncertainty measurement using the proposed evidential decision-making module. Then, it leverages the contextual discounted fusion layer to learn the reliability provided by each modality. Finally, a multi-level loss function is deployed for the optimization of evidence modeling and reliability learning. Moreover, this study elaborates on the framework interpretability by discussing the consistency between pixel attribution maps and the learned reliability coefficients. Extensive experiments are conducted on both multi-modality brain and liver datasets. The CDE-Net gains high performance with an average Dice score of 0.914 for brain tumor segmentation and 0.913 for liver tumor segmentation, which proves CDE-Net has great potential to facilitate the interpretation of artificial intelligence-based multi-modality medical image fusion.
RESUMO
With the ever-increasing number of artificial intelligence (AI) systems, mitigating risks associated with their use has become one of the most urgent scientific and societal issues. To this end, the European Union passed the EU AI Act, proposing solution strategies that can be summarized under the umbrella term trustworthiness. In anti-cancer drug sensitivity prediction, machine learning (ML) methods are developed for application in medical decision support systems, which require an extraordinary level of trustworthiness. This review offers an overview of the ML landscape of methods for anti-cancer drug sensitivity prediction, including a brief introduction to the four major ML realms (supervised, unsupervised, semi-supervised, and reinforcement learning). In particular, we address the question to what extent trustworthiness-related properties, more specifically, interpretability and reliability, have been incorporated into anti-cancer drug sensitivity prediction methods over the previous decade. In total, we analyzed 36 papers with approaches for anti-cancer drug sensitivity prediction. Our results indicate that the need for reliability has hardly been addressed so far. Interpretability, on the other hand, has often been considered for model development. However, the concept is rather used intuitively, lacking clear definitions. Thus, we propose an easily extensible taxonomy for interpretability, unifying all prevalent connotations explicitly or implicitly used within the field.
Assuntos
Antineoplásicos , Aprendizado de Máquina , Neoplasias , Humanos , Neoplasias/tratamento farmacológico , Antineoplásicos/uso terapêutico , Reprodutibilidade dos Testes , Inquéritos e Questionários , Resistencia a Medicamentos AntineoplásicosRESUMO
Safefy-critical domains often employ autonomous agents which follow a sequential decision-making setup, whereby the agent follows a policy to dictate the appropriate action at each step. AI-practitioners often employ reinforcement learning algorithms to allow an agent to find the best policy. However, sequential systems often lack clear and immediate signs of wrong actions, with consequences visible only in hindsight, making it difficult to humans to understand system failure. In reinforcement learning, this is referred to as the credit assignment problem. To effectively collaborate with an autonomous system, particularly in a safety-critical setting, explanations should enable a user to better understand the policy of the agent and predict system behavior so that users are cognizant of potential failures and these failures can be diagnosed and mitigated. However, humans are diverse and have innate biases or preferences which may enhance or impair the utility of a policy explanation of a sequential agent. Therefore, in this paper, we designed and conducted human-subjects experiment to identify the factors which influence the perceived usability with the objective usefulness of policy explanations for reinforcement learning agents in a sequential setting. Our study had two factors: the modality of policy explanation shown to the user (Tree, Text, Modified Text, and Programs) and the "first impression" of the agent, i.e., whether the user saw the agent succeed or fail in the introductory calibration video. Our findings characterize a preference-performance tradeoff wherein participants perceived language-based policy explanations to be significantly more useable; however, participants were better able to objectively predict the agent's behavior when provided an explanation in the form of a decision tree. Our results demonstrate that user-specific factors, such as computer science experience (p < 0.05), and situational factors, such as watching agent crash (p < 0.05), can significantly impact the perception and usefulness of the explanation. This research provides key insights to alleviate prevalent issues regarding innapropriate compliance and reliance, which are exponentially more detrimental in safety-critical settings, providing a path forward for XAI developers for future work on policy-explanations.
RESUMO
Ensuring the interpretability of machine learning models in chemical engineering remains challenging due to inherent limitations and data quality issues, hindering their reliable application. In this study, a qualitatively implicit knowledge-guided machine learning framework is proposed to improve plasma gasification modelling. Starting with a pre-trained machine learning model, parameters are further optimized by integrating the heuristic algorithm to minimize the data fitting errors and resolving implicit monotonic inconsistencies. The latter is comprehensively quantified through Monte Carlo simulations. This framework is adaptive to different machine learning techniques, exemplified by artificial neural network (ANN) and support vector machine (SVM) in this study. Validated by a case study on plasma gasification, the results reveal that the improved models achieve better generalizability and scientific interpretability in predicting syngas quality. Specifically, for ANN, the root mean square error (RMSE) and knowledge-based error (KE) reduce by 36.44% and 83.22%, respectively, while SVM displays a decrease of 2.58% in RMSE and a remarkable 100% in KE. Importantly, the improved models successfully capture all desired implicit monotonicity relationships between syngas quality and feedstock characteristics/operating parameters, addressing a limitation that traditional machine learning struggles with.
RESUMO
INTRODUCTION: Immune checkpoint inhibitors (ICIs) are potent and precise therapies for various cancer types, significantly improving survival rates in patients who respond positively to them. However, only a minority of patients benefit from ICI treatments. OBJECTIVES: Identifying ICI responders before treatment could greatly conserve medical resources, minimize potential drug side effects, and expedite the search for alternative therapies. Our goal is to introduce a novel deep-learning method to predict ICI treatment responses in cancer patients. METHODS: The proposed deep-learning framework leverages graph neural network and biological pathway knowledge. We trained and tested our method using ICI-treated patients' data from several clinical trials covering melanoma, gastric cancer, and bladder cancer. RESULTS: Our results demonstrate that this predictive model outperforms current state-of-the-art methods and tumor microenvironment-based predictors. Additionally, the model quantifies the importance of pathways, pathway interactions, and genes in its predictions. A web server for IRnet has been developed and deployed, providing broad accessibility to users at https://irnet.missouri.edu. CONCLUSION: IRnet is a competitive tool for predicting patient responses to immunotherapy, specifically ICIs. Its interpretability also offers valuable insights into the mechanisms underlying ICI treatments.
RESUMO
Explainable artificial intelligence (XAI) has gained significant attention in various domains, including natural and medical image analysis. However, its application in spectroscopy remains relatively unexplored. This systematic review aims to fill this gap by providing a comprehensive overview of the current landscape of XAI in spectroscopy and identifying potential benefits and challenges associated with its implementation. Following the PRISMA guideline 2020, we conducted a systematic search across major journal databases, resulting in 259 initial search results. After removing duplicates and applying inclusion and exclusion criteria, 21 scientific studies were included in this review. Notably, most of the studies focused on using XAI methods for spectral data analysis, emphasizing identifying significant spectral bands rather than specific intensity peaks. Among the most utilized AI techniques were SHapley Additive exPlanations (SHAP), masking methods inspired by Local Interpretable Model-agnostic Explanations (LIME), and Class Activation Mapping (CAM). These methods were favored due to their model-agnostic nature and ease of use, enabling interpretable explanations without modifying the original models. Future research should propose new methods and explore the adaptation of other XAI employed in other domains to better suit the unique characteristics of spectroscopic data.
RESUMO
Promoters are DNA sequences that bind with RNA polymerase to initiate transcription, regulating this process through interactions with transcription factors. Accurate identification of promoters is crucial for understanding gene expression regulation mechanisms and developing therapeutic approaches for various diseases. However, experimental techniques for promoter identification are often expensive, time-consuming, and inefficient, necessitating the development of accurate and efficient computational models for this task. Enhancing the model's ability to recognize promoters across multiple species and improving its interpretability pose significant challenges. In this study, we introduce a novel interpretable model based on graph neural networks, named GraphPro, for multi-species promoter identification. Initially, we encode the sequences using k-tuple nucleotide frequency pattern, dinucleotide physicochemical properties, and dna2vec. Subsequently, we construct two feature extraction modules based on convolutional neural networks and graph neural networks. These modules aim to extract specific motifs from the promoters, learn their dependencies, and capture the underlying structural features of the promoters, providing a more comprehensive representation. Finally, a fully connected neural network predicts whether the input sequence is a promoter. We conducted extensive experiments on promoter datasets from eight species, including Human, Mouse, and Escherichia coli. The experimental results show that the average Sn, Sp, Acc and MCC values of GraphPro are 0.9123, 0.9482, 0.8840 and 0.7984, respectively. Compared with previous promoter identification methods, GraphPro not only achieves better recognition accuracy on multiple species, but also outperforms all previous methods in cross-species prediction ability. Furthermore, by visualizing GraphPro's decision process and analyzing the sequences matching the transcription factor binding motifs captured by the model, we validate its significant advantages in biological interpretability. The source code for GraphPro is available at https://github.com/liuliwei1980/GraphPro.
RESUMO
The enhancement of fabric quality prediction in the textile manufacturing sector is achieved by utilizing information derived from sensors within the Internet of Things (IoT) and Enterprise Resource Planning (ERP) systems linked to sensors embedded in textile machinery. The integration of Industry 4.0 concepts is instrumental in harnessing IoT sensor data, which, in turn, leads to improvements in productivity and reduced lead times in textile manufacturing processes. This study addresses the issue of imbalanced data pertaining to fabric quality within the textile manufacturing industry. It encompasses an evaluation of seven open-source automated machine learning (AutoML) technologies, namely FLAML (Fast Lightweight AutoML), AutoViML (Automatically Build Variant Interpretable ML models), EvalML (Evaluation Machine Learning), AutoGluon, H2OAutoML, PyCaret, and TPOT (Tree-based Pipeline Optimization Tool). The most suitable solutions are chosen for certain circumstances by employing an innovative approach that finds a compromise among computational efficiency and forecast accuracy. The results reveal that EvalML emerges as the top-performing AutoML model for a predetermined objective function, particularly excelling in terms of mean absolute error (MAE). On the other hand, even with longer inference periods, AutoGluon performs better than other methods in measures like mean absolute percentage error (MAPE), root mean squared error (RMSE), and r-squared. Additionally, the study explores the feature importance rankings provided by each AutoML model, shedding light on the attributes that significantly influence predictive outcomes. Notably, sin/cos encoding is found to be particularly effective in characterizing categorical variables with a large number of unique values. This study includes useful information about the application of AutoML in the textile industry and provides a roadmap for employing Industry 4.0 technologies to enhance fabric quality prediction. The research highlights the importance of striking a balance between predictive accuracy and computational efficiency, emphasizes the significance of feature importance for model interpretability, and lays the groundwork for future investigations in this field.
RESUMO
Predicting the capacity of lithium-ion battery (LIB) plays a crucial role in ensuring the safe operation of LIBs and prolonging their lifespan. However, LIBs are easily affected by environmental interference, which may impact the precision of predictions. Furthermore, interpretability in the process of predicting LIB capacity is also important for users to understand the model, identify issues, and make decisions. In this study, an interpretable method considering environmental interference (IM-EI) for predicting LIB capacity is introduced. Spearman correlation coefficients, interpretability principles, belief rule base (BRB), and interpretability constraints are used to improve the prediction precision and interpretability of IM-EI. Dynamic attribute reliability is introduced to minimize the effect of environmental interference. The experimental results show that IM-EI model has good interpretability and high precision compared to the other models. Under interference conditions, the model still has good precision and robustness.
RESUMO
Preparation of activated carbons is an important way to utilize municipal sludge (MS) resources, while drying is a pretreatment method for making activated carbons from MS. In this study, machine learning techniques were used to develop moisture ratio (MR) and composting temperature (CT) prediction models for the thermally assisted biodrying process of MS. First, six machine learning (ML) models were used to construct the MR and CT prediction models, respectively. Then the hyperparameters of the ML models were optimized using the Bayesian optimization algorithm, and the prediction performances of these models after optimization were compared. Finally, the effect of each input feature on the model was also evaluated using SHapley Additive exPlanations (SHAP) analysis and Partial Dependence Plots (PDPs) analysis. The results showed that Gaussian process regression (GPR) was the best model for predicting MR and CT, with R2 of 0.9967 and 0.9958, respectively, and root mean square errors (RMSE) of 0.0059 and 0.354 â. In addition, graphical user interface software was developed to facilitate the use of the GPR model for predicting MR and CT by researchers and engineers. This study contributes to the rapid prediction, improvement, and optimization of MR and CT during thermally assisted biodrying of MS, and also provides valuable guidance for the dynamic regulation of the drying process.
RESUMO
Pediatric respiratory disease diagnosis and subsequent treatment require accurate and interpretable analysis. A chest X-ray is the most cost-effective and rapid method for identifying and monitoring various thoracic diseases in children. Recent developments in self-supervised and transfer learning have shown their potential in medical imaging, including chest X-ray areas. In this article, we propose a three-stage framework with knowledge transfer from adult chest X-rays to aid the diagnosis and interpretation of pediatric thorax diseases. We conducted comprehensive experiments with different pre-training and fine-tuning strategies to develop transformer or convolutional neural network models and then evaluate them qualitatively and quantitatively. The ViT-Base/16 model, fine-tuned with the CheXpert dataset, a large chest X-ray dataset, emerged as the most effective, achieving a mean AUC of 0.761 (95% CI: 0.759-0.763) across six disease categories and demonstrating a high sensitivity (average 0.639) and specificity (average 0.683), which are indicative of its strong discriminative ability. The baseline models, ViT-Small/16 and ViT-Base/16, when directly trained on the Pediatric CXR dataset, only achieved mean AUC scores of 0.646 (95% CI: 0.641-0.651) and 0.654 (95% CI: 0.648-0.660), respectively. Qualitatively, our model excels in localizing diseased regions, outperforming models pre-trained on ImageNet and other fine-tuning approaches, thus providing superior explanations. The source code is available online and the data can be obtained from PhysioNet.
RESUMO
Background: The application of machine learning in health care often necessitates the use of hierarchical codes such as the International Classification of Diseases (ICD) and Anatomical Therapeutic Chemical (ATC) systems. These codes classify diseases and medications, respectively, thereby forming extensive data dimensions. Unsupervised feature selection tackles the "curse of dimensionality" and helps to improve the accuracy and performance of supervised learning models by reducing the number of irrelevant or redundant features and avoiding overfitting. Techniques for unsupervised feature selection, such as filter, wrapper, and embedded methods, are implemented to select the most important features with the most intrinsic information. However, they face challenges due to the sheer volume of ICD and ATC codes and the hierarchical structures of these systems. Objective: The objective of this study was to compare several unsupervised feature selection methods for ICD and ATC code databases of patients with coronary artery disease in different aspects of performance and complexity and select the best set of features representing these patients. Methods: We compared several unsupervised feature selection methods for 2 ICD and 1 ATC code databases of 51,506 patients with coronary artery disease in Alberta, Canada. Specifically, we used the Laplacian score, unsupervised feature selection for multicluster data, autoencoder-inspired unsupervised feature selection, principal feature analysis, and concrete autoencoders with and without ICD or ATC tree weight adjustment to select the 100 best features from over 9000 ICD and 2000 ATC codes. We assessed the selected features based on their ability to reconstruct the initial feature space and predict 90-day mortality following discharge. We also compared the complexity of the selected features by mean code level in the ICD or ATC tree and the interpretability of the features in the mortality prediction task using Shapley analysis. Results: In feature space reconstruction and mortality prediction, the concrete autoencoder-based methods outperformed other techniques. Particularly, a weight-adjusted concrete autoencoder variant demonstrated improved reconstruction accuracy and significant predictive performance enhancement, confirmed by DeLong and McNemar tests (P<.05). Concrete autoencoders preferred more general codes, and they consistently reconstructed all features accurately. Additionally, features selected by weight-adjusted concrete autoencoders yielded higher Shapley values in mortality prediction than most alternatives. Conclusions: This study scrutinized 5 feature selection methods in ICD and ATC code data sets in an unsupervised context. Our findings underscore the superiority of the concrete autoencoder method in selecting salient features that represent the entire data set, offering a potential asset for subsequent machine learning research. We also present a novel weight adjustment approach for the concrete autoencoders specifically tailored for ICD and ATC code data sets to enhance the generalizability and interpretability of the selected features.
RESUMO
Amino acid scales are crucial for protein prediction tasks, many of them being curated in the AAindex database. Despite various clustering attempts to organize them and to better understand their relationships, these approaches lack the fine-grained classification necessary for satisfactory interpretability in many protein prediction problems. To address this issue, we developed AAontology-a two-level classification for 586 amino acid scales (mainly from AAindex) together with an in-depth analysis of their relations-using bag-of-word-based classification, clustering, and manual refinement over multiple iterations. AAontology organizes physicochemical scales into 8 categories and 67 subcategories, enhancing the interpretability of scale-based machine learning methods in protein bioinformatics. Thereby it enables researchers to gain a deeper biological insight. We anticipate that AAontology will be a building block to link amino acid properties with protein function and dysfunctions as well as aid informed decision-making in mutation analysis or protein drug design.
RESUMO
This review delves into the burgeoning field of explainable artificial intelligence (XAI) in the detection and analysis of lung diseases through vocal biomarkers. Lung diseases, often elusive in their early stages, pose a significant public health challenge. Recent advancements in AI have ushered in innovative methods for early detection, yet the black-box nature of many AI models limits their clinical applicability. XAI emerges as a pivotal tool, enhancing transparency and interpretability in AI-driven diagnostics. This review synthesizes current research on the application of XAI in analyzing vocal biomarkers for lung diseases, highlighting how these techniques elucidate the connections between specific vocal features and lung pathology. We critically examine the methodologies employed, the types of lung diseases studied, and the performance of various XAI models. The potential for XAI to aid in early detection, monitor disease progression, and personalize treatment strategies in pulmonary medicine is emphasized. Furthermore, this review identifies current challenges, including data heterogeneity and model generalizability, and proposes future directions for research. By offering a comprehensive analysis of explainable AI features in the context of lung disease detection, this review aims to bridge the gap between advanced computational approaches and clinical practice, paving the way for more transparent, reliable, and effective diagnostic tools.
Assuntos
Inteligência Artificial , Biomarcadores , Pneumopatias , Humanos , Pneumopatias/diagnóstico , Biomarcadores/metabolismoRESUMO
Respiratory diseases are one of the major health problems worldwide. Early diagnosis of the disease types is of vital importance. As one of the main symptoms of many respiratory diseases, cough may contain information about different pathological changes in the respiratory system. Therefore, many researchers have used cough sounds to diagnose different diseases through artificial intelligence in recent years. The acoustic features and data augmentation methods commonly used in speech tasks are used to achieve better performance. Although these methods are applicable, previous studies have not considered the characteristics of cough sound signals. In this paper, we designed a cough-based respiratory disease classification system and proposed audio characteristic-dependent feature extraction and data augmentation methods. Firstly, according to the short durations and rapid transition of different cough stages, we proposed maximum overlapping mel-spectrogram to avoid missing inter-frame information caused by traditional framing methods. Secondly, we applied various data augmentation methods to mitigate the problem of limited labeled data. Based on the frequency energy distributions of different diseased cough audios, we proposed a parameter-independent self-energy-based augmentation method to enhance the differences between different frequency bands. Finally, in the model testing stage, we leveraged test-time augmentation to further improve the classification performance by fusing the test results of the original and multiple augmented audios. The proposed methods were validated on the Coswara dataset through stratified four-fold cross-validation. Compared to the baseline model using mel-spectrogram as input, the proposed methods achieved an average absolute performance improvement of 3.33% and 3.10% in macro Area Under the Receiver Operating Characteristic (macro AUC) and Unweighted Average Recall (UAR), respectively. The visualization results through Gradient-weighted Class Activation Mapping (Grad-CAM) showed the contributions of different features to model decisions.
Assuntos
Tosse , Humanos , Tosse/classificação , Tosse/fisiopatologia , Processamento de Sinais Assistido por Computador , Masculino , Feminino , Espectrografia do Som/métodos , Adulto , Pessoa de Meia-IdadeRESUMO
Machine learning models are often viewed as black boxes in landslide susceptibility assessment, lacking an analysis of how input features predict outcomes. This makes it challenging to understand the mechanisms and key factors behind landslides. To enhance the interpretability of machine learning models in wide-area landslide susceptibility assessments, this study uses the Shapely method to explore the contributions of feature factors from local, global, and spatial perspectives. Landslide susceptibility assessments were conducted using random forest (RF), support vector machine (SVM), and eXtreme Gradient Boosting (XGBoost) models, focusing on the geologically complex Sichuan-Tibet region. Initially, the study revealed the contributions of specific key feature factors to landslides from a local perspective. It then examines the overall impact of interactions among feature factors on landslide occurrence globally. Finally, it unveils the spatial distribution patterns of the contributions of various feature factors to landslide occurrence. The analysis indicates the following: (1) The XGBoost model excels in landslide susceptibility assessment, achieving accuracy, precision, recall, F1-score, and AUC values of 0.7815, 0.7858, 0.7962, 0.7910, and 0.86, respectively; (2) The Shapely method identifies the leading factors for landslides in the Sichuan-Tibet region as Elevation (3000-4000 m), PGA (1-2 g), NDVI (<0.5), and distance to rivers (<3 km); (3) Using the Shapely method, the study explains the contributions, interaction mechanisms, and spatial distribution patterns of landslide susceptibility feature factors across local, global, and spatial perspectives. These findings offer new avenues and methods for the in-depth exploration and scientific prediction of landslide risks.
Assuntos
Deslizamentos de Terra , Tibet , Aprendizado de Máquina , Máquina de Vetores de Suporte , ChinaRESUMO
Establishing the relationships among hierarchical visual attributes of objects in the visual world is crucial for human cognition. The classic convolution neural network (CNN) can successfully extract hierarchical features but ignore the relationships among features, resulting in shortcomings compared to humans in areas like interpretability and domain generalization. Recently, algorithms have introduced feature relationships by external prior knowledge and special auxiliary modules, which have been proven to bring multiple improvements in many computer vision tasks. However, prior knowledge is often difficult to obtain, and auxiliary modules bring additional consumption of computing and storage resources, which limits the flexibility and practicality of the algorithm. In this paper, we aim to drive the CNN model to learn the relationships among hierarchical deep features without prior knowledge and consumption increasing, while enhancing the fundamental performance of some aspects. Firstly, the task of learning the relationships among hierarchical features in CNN is defined and three key problems related to this task are pointed out, including the quantitative metric of connection intensity, the threshold of useless connections, and the updating strategy of relation graph. Secondly, Relational Embedding Convolution (RE-Conv) layer is proposed for the representation of feature relationships in convolution layer, followed by a scheme called use & disuse strategy which aims to address the three problems of feature relation learning. Finally, the improvements brought by the proposed feature relation learning scheme have been demonstrated through numerous experiments, including interpretability, domain generalization, noise robustness, and inference efficiency. In particular, the proposed scheme outperforms many state-of-the-art methods in the domain generalization community and can be seamlessly integrated with existing methods for further improvement. Meanwhile, it maintains comparable precision to the original CNN model while reducing floating point operations (FLOPs) by approximately 50%.
RESUMO
A better understanding of the emergent computation and problem-solving capabilities of recent large language models is of paramount importance to further improve them and broaden their applicability. This work investigates how a language model, trained to predict the next token, can perform arithmetic computations generalizing beyond training data. Binary addition and multiplication constitute a good testbed for this purpose, since they require a very small vocabulary and exhibit relevant input/output discontinuities making smooth input interpolation ineffective for novel data. We successfully trained a light language model to learn these tasks and ran a number of experiments to investigate the extrapolation capabilities and internal information processing. Our findings support the hypothesis that the language model works as an Encoding-Regression-Decoding machine where the computation takes place in the value space once the input token representation is mapped to an appropriate internal representation.
RESUMO
Aims: Deep learning methods have recently gained success in detecting left ventricular systolic dysfunction (LVSD) from electrocardiogram (ECG) waveforms. Despite their high level of accuracy, they are difficult to interpret and deploy broadly in the clinical setting. In this study, we set out to determine whether simpler models based on standard ECG measurements could detect LVSD with similar accuracy to that of deep learning models. Methods and results: Using an observational data set of 40 994 matched 12-lead ECGs and transthoracic echocardiograms, we trained a range of models with increasing complexity to detect LVSD based on ECG waveforms and derived measurements. The training data were acquired from the Stanford University Medical Center. External validation data were acquired from the Columbia Medical Center and the UK Biobank. The Stanford data set consisted of 40 994 matched ECGs and echocardiograms, of which 9.72% had LVSD. A random forest model using 555 discrete, automated measurements achieved an area under the receiver operator characteristic curve (AUC) of 0.92 (0.91-0.93), similar to a deep learning waveform model with an AUC of 0.94 (0.93-0.94). A logistic regression model based on five measurements achieved high performance [AUC of 0.86 (0.85-0.87)], close to a deep learning model and better than N-terminal prohormone brain natriuretic peptide (NT-proBNP). Finally, we found that simpler models were more portable across sites, with experiments at two independent, external sites. Conclusion: Our study demonstrates the value of simple electrocardiographic models that perform nearly as well as deep learning models, while being much easier to implement and interpret.
RESUMO
Computer-aided diagnosis (CAD) for thyroid nodules has been studied for years, yet there are still reliability and interpretability challenges due to the lack of clinically-relevant evidence. To address this issue, inspired by Thyroid Imaging Reporting and Data System (TI-RADS), we propose a novel interpretable two-branch bi-coordinate network based on multi-grained domain knowledge. First, we transform the two types of domain knowledge provided by TI-RADS, namely region-based and boundary-based knowledge, into labels at multi-grained levels: coarse-grained classification labels, and fine-grained region segmentation masks and boundary localization vectors. We combine these two labels to form the Multi-grained Domain Knowledge Representation (MG-DKR) of TI-RADS. Then we design a Two-branch Bi-coordinate network (TB2C-net) which utilizes two branches to predict MG-DKR from both Cartesian and polar images, and uses an attention-based integration module to integrate the features of the two branches for benign-malignant classification. We validated our method on a large cohort containing 3245 patients (with 3558 nodules and 6466 ultrasound images). Results show that our method achieves competitive performance with AUC of 0.93 and ACC of 0.87 compared with other state-of-the-art methods. Ablation experiment results demonstrate the effectiveness of the TB2C-net and MG-DKR, and the knowledge attention map from the integration module provides the interpretability for benign-malignant classification.