INSURE: An Information Theory iNspired diSentanglement and pURification modEl for Domain Generalization.

Domain Generalization (DG) aims to learn a generalizable model on the unseen target domain by only training on the multiple observed source domains. Although a variety of DG methods have focused on extracting domain-invariant features, the domain-specific class-relevant features have attracted attention and been argued to benefit generalization to the unseen target domain. To take into account the class-relevant domain-specific information, in this paper we propose an Information theory iNspired diSentanglement and pURification modEl (INSURE) to explicitly disentangle the latent features to obtain sufficient and compact (necessary) class-relevant feature for generalization to the unseen domain. Specifically, we first propose an information theory inspired loss function to ensure the disentangled class-relevant features contain sufficient class label information and the other disentangled auxiliary feature has sufficient domain information. We further propose a paired purification loss function to let the auxiliary feature discard all the class-relevant information and thus the class-relevant feature will contain sufficient and compact (necessary) class-relevant information. Moreover, instead of using multiple encoders, we propose to use a learnable binary mask as our disentangler to make the disentanglement more efficient and make the disentangled features complementary to each other. We conduct extensive experiments on five widely used DG benchmark datasets including PACS, VLCS, OfficeHome, TerraIncognita, and DomainNet. The proposed INSURE achieves state-of-the-art performance. We also empirically show that domain-specific class-relevant features are beneficial for domain generalization. The code is available at https://github.com/yuxi120407/INSURE.

Toward Real-World Voice Disorder Classification.

OBJECTIVE: Voice disorders significantly compromise individuals' ability to speak in their daily lives. Without early diagnosis and treatment, these disorders may deteriorate drastically. Thus, automatic classification systems at home are desirable for people who are inaccessible to clinical disease assessments. However, the performance of such systems may be weakened due to the constrained resources and domain mismatch between the clinical data and noisy real-world data. METHODS: This study develops a compact and domain-robust voice disorder classification system to identify the utterances of health, neoplasm, and benign structural diseases. Our proposed system utilizes a feature extractor model composed of factorized convolutional neural networks and subsequently deploys domain adversarial training to reconcile the domain mismatch by extracting domain-invariant features. RESULTS: The results show that the unweighted average recall in the noisy real-world domain improved by 13% and remained at 80% in the clinic domain with only slight degradation. The domain mismatch was effectively eliminated. Moreover, the proposed system reduced the usage of both memory and computation by over 73.9%. CONCLUSION: By deploying factorized convolutional neural networks and domain adversarial training, domain-invariant features can be derived for voice disorder classification with limited resources. The promising results confirm that the proposed system can significantly reduce resource consumption and improve classification accuracy by considering the domain mismatch. SIGNIFICANCE: To the best of our knowledge, this is the first study that jointly considers real-world model compression and noise-robustness issues in voice disorder classification. The proposed system is intended for application to embedded systems with limited resources.

A situational awareness Bayesian network approach for accurate and credible personalized adaptive radiotherapy outcomes prediction in lung cancer patients.

PURPOSE: A situational awareness Bayesian network (SA-BN) approach is developed to improve physicians' trust in the prediction of radiation outcomes and evaluate its performance for personalized adaptive radiotherapy (pART). METHODS: 118 non-small-cell lung cancer patients with their biophysical features were employed for discovery (n = 68) and validation (n = 50) of radiation outcomes prediction modeling. Patients' important characteristics identified by radiation experts to predict individual's tumor local control (LC) or radiation pneumonitis with grade ≥ 2 (RP2) were incorporated as expert knowledge (EK). Besides generating an EK-based naïve BN (EK-NBN), an SA-BN was developed by incorporating the EK features into pure data-driven BN (PD-BN) methods to improve the credibility of LC or / and RP2 prediction. After using area under the free-response receiver operating characteristics curve (AU-FROC) to assess the joint prediction of these outcomes, their prediction performances were compared with a regression approach based on the expert yielded estimates (EYE) penalty and its variants. RESULTS: In addition to improving the credibility of radiation outcomes prediction, the SA-BN approach outperformed the EYE penalty and its variants in terms of the joint prediction of LC and RP2. The value of AU-FROC improves from 0.70 (95% CI: 0.54-0.76) using EK-NBN, to 0.75 (0.65-0.82) using a variant of EYE penalty, to 0.83 (0.75-0.93) using PD-BN and 0.83 (0.77-0.90) using SA-BN; with similar trends in the validation cohort. CONCLUSIONS: The SA-BN approach can provide an accurate and credible human-machine interface to gain physicians' trust in clinical decision-making, which has the potential to be an important component of pART.

Introduction to machine and deep learning for medical physicists.

Recent years have witnessed tremendous growth in the application of machine learning (ML) and deep learning (DL) techniques in medical physics. Embracing the current big data era, medical physicists equipped with these state-of-the-art tools should be able to solve pressing problems in modern radiation oncology. Here, a review of the basic aspects involved in ML/DL model building, including data processing, model training, and validation for medical physics applications is presented and discussed. Machine learning can be categorized based on the underlying task into supervised learning, unsupervised learning, or reinforcement learning; each of these categories has its own input/output dataset characteristics and aims to solve different classes of problems in medical physics ranging from automation of processes to predictive analytics. It is recognized that data size requirements may vary depending on the specific medical physics application and the nature of the algorithms applied. Data processing, which is a crucial step for model stability and precision, should be performed before training the model. Deep learning as a subset of ML is able to learn multilevel representations from raw input data, eliminating the necessity for hand crafted features in classical ML. It can be thought of as an extension of the classical linear models but with multilayer (deep) structures and nonlinear activation functions. The logic of going "deeper" is related to learning complex data structures and its realization has been aided by recent advancements in parallel computing architectures and the development of more robust optimization methods for efficient training of these algorithms. Model validation is an essential part of ML/DL model building. Without it, the model being developed cannot be easily trusted to generalize to unseen data. Whenever applying ML/DL, one should keep in mind, according to Amara's law, that humans may tend to overestimate the ability of a technology in the short term and underestimate its capability in the long term. To establish ML/DL role into standard clinical workflow, models considering balance between accuracy and interpretability should be developed. Machine learning/DL algorithms have potential in numerous radiation oncology applications, including automatizing mundane procedures, improving efficiency and safety of auto-contouring, treatment planning, quality assurance, motion management, and outcome predictions. Medical physicists have been at the frontiers of technology translation into medicine and they ought to be prepared to embrace the inevitable role of ML/DL in the practice of radiation oncology and lead its clinical implementation.

Machine Learning and Imaging Informatics in Oncology.

In the era of personalized and precision medicine, informatics technologies utilizing machine learning (ML) and quantitative imaging are witnessing a rapidly increasing role in medicine in general and in oncology in particular. This expanding role ranges from computer-aided diagnosis to decision support of treatments with the potential to transform the current landscape of cancer management. In this review, we aim to provide an overview of ML methodologies and imaging informatics techniques and their recent application in modern oncology. We will review example applications of ML in oncology from the literature, identify current challenges and highlight future potentials.

Quantum-inspired algorithm for radiotherapy planning optimization.

PURPOSE: Modern inverse radiotherapy treatment planning requires nonconvex, large-scale optimizations that must be solved within a clinically feasible timeframe. We have developed and tested a quantum-inspired, stochastic algorithm for intensity-modulated radiotherapy (IMRT): quantum tunnel annealing (QTA). By modeling the likelihood probability of accepting a higher energy solution after a particle tunneling through a potential energy barrier, QTA features an additional degree of freedom (the barrier width, w) not shared by traditional stochastic optimization methods such as Simulated Annealing (SA). This additional degree of freedom can improve convergence rates and achieve a more efficient and, potentially, effective treatment planning process. METHODS: To analyze the character of the proposed QTA algorithm, we chose two stereotactic body radiation therapy (SBRT) liver cases of variable complexity. The "easy" first case was used to confirm functionality, while the second case, with a more challenging geometry, was used to characterize and evaluate the QTA algorithm performance. Plan quality was assessed using dose-volume histogram-based objectives and dose distributions. Due to the stochastic nature of the solution search space, extensive tests were also conducted to determine the optimal smoothing technique, ensuring balance between plan deliverability and the resulting plan quality. QTA convergence rates were investigated in relation to the chosen barrier width function, and QTA and SA performances were compared regarding sensitivity to the choice of solution initializations, annealing schedules, and complexity of the dose-volume constraints. Finally, we investigated the extension from beamlet intensity optimization to direct aperture optimization (DAO). Influence matrices were calculated using the Eclipse scripting application program interface (API), and the optimizations were run on the University of Michigan's high-performance computing cluster, Flux. RESULTS: Our results indicate that QTA's barrier-width function can be tuned to achieve faster convergence rates. The QTA algorithm reached convergence up to 46.6% faster than SA for beamlet intensity optimization and up to 26.8% faster for DAO. QTA and SA were ultimately found to be equally insensitive to the initialization process, but the convergence rate of QTA was found to be more sensitive to the complexity of the dose-volume constraints. The optimal smoothing technique was found to be a combination of a Laplace-of-Gaussian (LOG) edge-finding filter implemented as a penalty within the objective function and a two-dimensional Savitzky-Golay filter applied to the final iteration; this achieved total monitor units more than 20% smaller than plans optimized by commercial treatment planning software. CONCLUSIONS: We have characterized the performance of a stochastic, quantum-inspired optimization algorithm, QTA, for radiotherapy treatment planning. This proof of concept study suggests that QTA can be tuned to achieve faster convergence than SA; therefore, QTA may be a good candidate for future knowledge-based or adaptive radiation therapy applications.

Combining handcrafted features with latent variables in machine learning for prediction of radiation-induced lung damage.

PURPOSE: There has been burgeoning interest in applying machine learning methods for predicting radiotherapy outcomes. However, the imbalanced ratio of a large number of variables to a limited sample size in radiation oncology constitutes a major challenge. Therefore, dimensionality reduction methods can be a key to success. The study investigates and contrasts the application of traditional machine learning methods and deep learning approaches for outcome modeling in radiotherapy. In particular, new joint architectures based on variational autoencoder (VAE) for dimensionality reduction are presented and their application is demonstrated for the prediction of lung radiation pneumonitis (RP) from a large-scale heterogeneous dataset. METHODS: A large-scale heterogeneous dataset containing a pool of 230 variables including clinical factors (e.g., dose, KPS, stage) and biomarkers (e.g., single nucleotide polymorphisms (SNPs), cytokines, and micro-RNAs) in a population of 106 nonsmall cell lung cancer (NSCLC) patients who received radiotherapy was used for modeling RP. Twenty-two patients had grade 2 or higher RP. Four methods were investigated, including feature selection (case A) and feature extraction (case B) with traditional machine learning methods, a VAE-MLP joint architecture (case C) with deep learning and lastly, the combination of feature selection and joint architecture (case D). For feature selection, Random forest (RF), Support Vector Machine (SVM), and multilayer perceptron (MLP) were implemented to select relevant features. Specifically, each method was run for multiple times to rank features within several cross-validated (CV) resampled sets. A collection of ranking lists were then aggregated by top 5% and Kemeny graph methods to identify the final ranking for prediction. A synthetic minority oversampling technique was applied to correct for class imbalance during this process. For deep learning, a VAE-MLP joint architecture where a VAE aimed for dimensionality reduction and an MLP aimed for classification was developed. In this architecture, reconstruction loss and prediction loss were combined into a single loss function to realize simultaneous training and weights were assigned to different classes to mitigate class imbalance. To evaluate the prediction performance and conduct comparisons, the area under receiver operating characteristic curves (AUCs) were performed for nested CVs for both handcrafted feature selections and the deep learning approach. The significance of differences in AUCs was assessed using the DeLong test of U-statistics. RESULTS: An MLP-based method using weight pruning (WP) feature selection yielded the best performance among the different hand-crafted feature selection methods (case A), reaching an AUC of 0.804 (95% CI: 0.761-0.823) with 29 top features. A VAE-MLP joint architecture (case C) achieved a comparable but slightly lower AUC of 0.781 (95% CI: 0.737-0.808) with the size of latent dimension being 2. The combination of handcrafted features (case A) and latent representation (case D) achieved a significant AUC improvement of 0.831 (95% CI: 0.805-0.863) with 22 features (P-value = 0.000642 compared with handcrafted features only (Case A) and P-value = 0.000453 compared to VAE alone (Case C)) with an MLP classifier. CONCLUSION: The potential for combination of traditional machine learning methods and deep learning VAE techniques has been demonstrated for dealing with limited datasets in modeling radiotherapy toxicities. Specifically, latent variables from a VAE-MLP joint architecture are able to complement handcrafted features for the prediction of RP and improve prediction over either method alone.

Balancing accuracy and interpretability of machine learning approaches for radiation treatment outcomes modeling.

Radiation outcomes prediction (ROP) plays an important role in personalized prescription and adaptive radiotherapy. A clinical decision may not only depend on an accurate radiation outcomes' prediction, but also needs to be made based on an informed understanding of the relationship among patients' characteristics, radiation response and treatment plans. As more patients' biophysical information become available, machine learning (ML) techniques will have a great potential for improving ROP. Creating explainable ML methods is an ultimate task for clinical practice but remains a challenging one. Towards complete explainability, the interpretability of ML approaches needs to be first explored. Hence, this review focuses on the application of ML techniques for clinical adoption in radiation oncology by balancing accuracy with interpretability of the predictive model of interest. An ML algorithm can be generally classified into an interpretable (IP) or non-interpretable (NIP) ("black box") technique. While the former may provide a clearer explanation to aid clinical decision-making, its prediction performance is generally outperformed by the latter. Therefore, great efforts and resources have been dedicated towards balancing the accuracy and the interpretability of ML approaches in ROP, but more still needs to be done. In this review, current progress to increase the accuracy for IP ML approaches is introduced, and major trends to improve the interpretability and alleviate the "black box" stigma of ML in radiation outcomes modeling are summarized. Efforts to integrate IP and NIP ML approaches to produce predictive models with higher accuracy and interpretability for ROP are also discussed.

The Role of Machine Learning in Knowledge-Based Response-Adapted Radiotherapy.

With the continuous increase in radiotherapy patient-specific data from multimodality imaging and biotechnology molecular sources, knowledge-based response-adapted radiotherapy (KBR-ART) is emerging as a vital area for radiation oncology personalized treatment. In KBR-ART, planned dose distributions can be modified based on observed cues in patients' clinical, geometric, and physiological parameters. In this paper, we present current developments in the field of adaptive radiotherapy (ART), the progression toward KBR-ART, and examine several applications of static and dynamic machine learning approaches for realizing the KBR-ART framework potentials in maximizing tumor control and minimizing side effects with respect to individual radiotherapy patients. Specifically, three questions required for the realization of KBR-ART are addressed: (1) what knowledge is needed; (2) how to estimate RT outcomes accurately; and (3) how to adapt optimally. Different machine learning algorithms for KBR-ART application shall be discussed and contrasted. Representative examples of different KBR-ART stages are also visited.

Radiomics in precision medicine for lung cancer.

With the improvement of external radiotherapy delivery accuracy, such as intensity-modulated and stereotactic body radiation therapy, radiation oncology has recently entered in the era of precision medicine. Despite these precise irradiation modalities, lung cancers remain one of the most aggressive human cancers worldwide, possibly because of diverse genotypic alterations that drive and maintain lung tumorigenesis. It has been long recognized that imaging could aid in the diagnosis, tumor delineation, and monitoring of lung cancer. Moreover, accumulating evidence suggests that imaging information could be further used to tailor treatment type and intensity, as well as predict treatment outcomes in radiotherapy. However, these imaging tasks have been carried out either qualitatively or using simplistic metrics that doesn't take advantage of the full scale of imaging knowledge. Radiomics, which is a recent field of research that aims to provide a more quantitative representation of imaging information relating tumor phenotypes to clinical and genotypic endpoints by embedding extracted image features into predictive mathematical models. These predictive models can be a key component in the clinician decision making and treatment personalization. This review provides an overview of the radiomics application and its methodology for radiation oncology studies in lung cancer.

Deep reinforcement learning for automated radiation adaptation in lung cancer.

PURPOSE: To investigate deep reinforcement learning (DRL) based on historical treatment plans for developing automated radiation adaptation protocols for nonsmall cell lung cancer (NSCLC) patients that aim to maximize tumor local control at reduced rates of radiation pneumonitis grade 2 (RP2). METHODS: In a retrospective population of 114 NSCLC patients who received radiotherapy, a three-component neural networks framework was developed for deep reinforcement learning (DRL) of dose fractionation adaptation. Large-scale patient characteristics included clinical, genetic, and imaging radiomics features in addition to tumor and lung dosimetric variables. First, a generative adversarial network (GAN) was employed to learn patient population characteristics necessary for DRL training from a relatively limited sample size. Second, a radiotherapy artificial environment (RAE) was reconstructed by a deep neural network (DNN) utilizing both original and synthetic data (by GAN) to estimate the transition probabilities for adaptation of personalized radiotherapy patients' treatment courses. Third, a deep Q-network (DQN) was applied to the RAE for choosing the optimal dose in a response-adapted treatment setting. This multicomponent reinforcement learning approach was benchmarked against real clinical decisions that were applied in an adaptive dose escalation clinical protocol. In which, 34 patients were treated based on avid PET signal in the tumor and constrained by a 17.2% normal tissue complication probability (NTCP) limit for RP2. The uncomplicated cure probability (P+) was used as a baseline reward function in the DRL. RESULTS: Taking our adaptive dose escalation protocol as a blueprint for the proposed DRL (GAN + RAE + DQN) architecture, we obtained an automated dose adaptation estimate for use at ∼2/3 of the way into the radiotherapy treatment course. By letting the DQN component freely control the estimated adaptive dose per fraction (ranging from 1-5 Gy), the DRL automatically favored dose escalation/de-escalation between 1.5 and 3.8 Gy, a range similar to that used in the clinical protocol. The same DQN yielded two patterns of dose escalation for the 34 test patients, but with different reward variants. First, using the baseline P+ reward function, individual adaptive fraction doses of the DQN had similar tendencies to the clinical data with an RMSE = 0.76 Gy; but adaptations suggested by the DQN were generally lower in magnitude (less aggressive). Second, by adjusting the P+ reward function with higher emphasis on mitigating local failure, better matching of doses between the DQN and the clinical protocol was achieved with an RMSE = 0.5 Gy. Moreover, the decisions selected by the DQN seemed to have better concordance with patients eventual outcomes. In comparison, the traditional temporal difference (TD) algorithm for reinforcement learning yielded an RMSE = 3.3 Gy due to numerical instabilities and lack of sufficient learning. CONCLUSION: We demonstrated that automated dose adaptation by DRL is a feasible and a promising approach for achieving similar results to those chosen by clinicians. The process may require customization of the reward function if individual cases were to be considered. However, development of this framework into a fully credible autonomous system for clinical decision support would require further validation on larger multi-institutional datasets.

Should regional ventilation function be considered during radiation treatment planning to prevent radiation-induced complications?

PURPOSE: To investigate the incorporation of pretherapy regional ventilation function in predicting radiation fibrosis (RF) in stage III nonsmall cell lung cancer (NSCLC) patients treated with concurrent thoracic chemoradiotherapy. METHODS: Thirty-seven patients with stage III NSCLC were retrospectively studied. Patients received one cycle of cisplatin-gemcitabine, followed by two to three cycles of cisplatin-etoposide concurrently with involved-field thoracic radiotherapy (46-66 Gy; 2 Gy/fraction). Pretherapy regional ventilation images of the lung were derived from 4D computed tomography via a density change-based algorithm with mass correction. In addition to the conventional dose-volume metrics (V20, V30, V40, and mean lung dose), dose-function metrics (fV20, fV30, fV40, and functional mean lung dose) were generated by combining regional ventilation and radiation dose. A new class of metrics was derived and referred to as dose-subvolume metrics (sV20, sV30, sV40, and subvolume mean lung dose); these were defined as the conventional dose-volume metrics computed on the functional lung. Area under the receiver operating characteristic curve (AUC) values and logistic regression analyses were used to evaluate these metrics in predicting hallmark characteristics of RF (lung consolidation, volume loss, and airway dilation). RESULTS: AUC values for the dose-volume metrics in predicting lung consolidation, volume loss, and airway dilation were 0.65-0.69, 0.57-0.70, and 0.69-0.76, respectively. The respective ranges for dose-function metrics were 0.63-0.66, 0.61-0.71, and 0.72-0.80 and for dose-subvolume metrics were 0.50-0.65, 0.65-0.75, and 0.73-0.85. Using an AUC value = 0.70 as cutoff value suggested that at least one of each type of metrics (dose-volume, dose-function, dose-subvolume) was predictive for volume loss and airway dilation, whereas lung consolidation cannot be accurately predicted by any of the metrics. Logistic regression analyses showed that dose-function and dose-subvolume metrics were significant (P values ≤ 0.02) in predicting volume airway dilation. Likelihood ratio test showed that when combining dose-function and/or dose-subvolume metrics with dose-volume metrics, the achieved improvements of prediction accuracy on volume loss and airway dilation were significant (P values ≤ 0.04). CONCLUSIONS: The authors' results demonstrated that the inclusion of regional ventilation function improved accuracy in predicting RF. In particular, dose-subvolume metrics provided a promising method for preventing radiation-induced pulmonary complications.