Data Characterization for Reliable AI in Medicine.

Research in Artificial Intelligence (AI)-based medical computer vision algorithms bear promises to improve disease screening, diagnosis, and subsequently patient care. However, these algorithms are highly impacted by the characteristics of the underlying data. In this work, we discuss various data characteristics, namely Volume, Veracity, Validity, Variety, and Velocity, that impact the design, reliability, and evolution of machine learning in medical computer vision. Further, we discuss each characteristic and the recent works conducted in our research lab that informed our understanding of the impact of these characteristics on the design of medical decision-making algorithms and outcome reliability.

Assessing Inter-Annotator Agreement for Medical Image Segmentation.

Artificial Intelligence (AI)-based medical computer vision algorithm training and evaluations depend on annotations and labeling. However, variability between expert annotators introduces noise in training data that can adversely impact the performance of AI algorithms. This study aims to assess, illustrate and interpret the inter-annotator agreement among multiple expert annotators when segmenting the same lesion(s)/abnormalities on medical images. We propose the use of three metrics for the qualitative and quantitative assessment of inter-annotator agreement: 1) use of a common agreement heatmap and a ranking agreement heatmap; 2) use of the extended Cohen's kappa and Fleiss' kappa coefficients for a quantitative evaluation and interpretation of inter-annotator reliability; and 3) use of the Simultaneous Truth and Performance Level Estimation (STAPLE) algorithm, as a parallel step, to generate ground truth for training AI models and compute Intersection over Union (IoU), sensitivity, and specificity to assess the inter-annotator reliability and variability. Experiments are performed on two datasets, namely cervical colposcopy images from 30 patients and chest X-ray images from 336 tuberculosis (TB) patients, to demonstrate the consistency of inter-annotator reliability assessment and the importance of combining different metrics to avoid bias assessment.

Unsupervised Deep Learning Registration of Uterine Cervix Sequence Images.

During a colposcopic examination of the uterine cervix for cervical cancer prevention, one or more digital images are typically acquired after the application of diluted acetic acid. An alternative approach is to acquire a sequence of images at fixed intervals during an examination before and after applying acetic acid. This approach is asserted to be more informative as it can capture dynamic pixel intensity variations on the cervical epithelium during the aceto-whitening reaction. However, the resulting time sequence images may not be spatially aligned due to the movement of the cervix with respect to the imaging device. Disease prediction using automated visual evaluation (AVE) techniques using multiple images could be adversely impacted without correction for this misalignment. The challenge is that there is no registration ground truth to help train a supervised-learning-based image registration algorithm. We present a novel unsupervised registration approach to align a sequence of digital cervix color images. The proposed deep-learning-based registration network consists of three branches and processes the red, green, and blue (RGB, respectively) channels of each input color image separately using an unsupervised strategy. Each network branch consists of a convolutional neural network (CNN) unit and a spatial transform unit. To evaluate the registration performance on a dataset that has no ground truth, we propose an evaluation strategy that is based on comparing automatic cervix segmentation masks in the registered sequence and the original sequence. The compared segmentation masks are generated by a fine-tuned transformer-based object detection model (DeTr). The segmentation model achieved Dice/IoU scores of 0.917/0.870 and 0.938/0.885, which are comparable to the performance of our previous model in two datasets. By comparing our segmentation on both original and registered time sequence images, we observed an average improvement in Dice scores of 12.62% following registration. Further, our approach achieved higher Dice and IoU scores and maintained full image integrity compared to a non-deep learning registration method on the same dataset.

A Deep Modality-Specific Ensemble for Improving Pneumonia Detection in Chest X-rays.

Pneumonia is an acute respiratory infectious disease caused by bacteria, fungi, or viruses. Fluid-filled lungs due to the disease result in painful breathing difficulties and reduced oxygen intake. Effective diagnosis is critical for appropriate and timely treatment and improving survival. Chest X-rays (CXRs) are routinely used to screen for the infection. Computer-aided detection methods using conventional deep learning (DL) models for identifying pneumonia-consistent manifestations in CXRs have demonstrated superiority over traditional machine learning approaches. However, their performance is still inadequate to aid in clinical decision-making. This study improves upon the state of the art as follows. Specifically, we train a DL classifier on large collections of CXR images to develop a CXR modality-specific model. Next, we use this model as the classifier backbone in the RetinaNet object detection network. We also initialize this backbone using random weights and ImageNet-pretrained weights. Finally, we construct an ensemble of the best-performing models resulting in improved detection of pneumonia-consistent findings. Experimental results demonstrate that an ensemble of the top-3 performing RetinaNet models outperformed individual models in terms of the mean average precision (mAP) metric (0.3272, 95% CI: (0.3006,0.3538)) toward this task, which is markedly higher than the state of the art (mAP: 0.2547). This performance improvement is attributed to the key modifications in initializing the weights of classifier backbones and constructing model ensembles to reduce prediction variance compared to individual constituent models.

A Systematic Evaluation of Ensemble Learning Methods for Fine-Grained Semantic Segmentation of Tuberculosis-Consistent Lesions in Chest Radiographs.

Automated segmentation of tuberculosis (TB)-consistent lesions in chest X-rays (CXRs) using deep learning (DL) methods can help reduce radiologist effort, supplement clinical decision-making, and potentially result in improved patient treatment. The majority of works in the literature discuss training automatic segmentation models using coarse bounding box annotations. However, the granularity of the bounding box annotation could result in the inclusion of a considerable fraction of false positives and negatives at the pixel level that may adversely impact overall semantic segmentation performance. This study evaluates the benefits of using fine-grained annotations of TB-consistent lesions toward training the variants of U-Net models and constructing their ensembles for semantically segmenting TB-consistent lesions in both original and bone-suppressed frontal CXRs. The segmentation performance is evaluated using several ensemble methods such as bitwise- AND, bitwise-OR, bitwise-MAX, and stacking. Extensive empirical evaluations showcased that the stacking ensemble demonstrated superior segmentation performance (Dice score: 0.5743, 95% confidence interval: (0.4055, 0.7431)) compared to the individual constituent models and other ensemble methods. To the best of our knowledge, this is the first study to apply ensemble learning to improve fine-grained TB-consistent lesion segmentation performance.

Uncertainty Quantification in Segmenting Tuberculosis-Consistent Findings in Frontal Chest X-rays.

Deep learning (DL) methods have demonstrated superior performance in medical image segmentation tasks. However, selecting a loss function that conforms to the data characteristics is critical for optimal performance. Further, the direct use of traditional DL models does not provide a measure of uncertainty in predictions. Even high-quality automated predictions for medical diagnostic applications demand uncertainty quantification to gain user trust. In this study, we aim to investigate the benefits of (i) selecting an appropriate loss function and (ii) quantifying uncertainty in predictions using a VGG16-based-U-Net model with the Monto-Carlo (MCD) Dropout method for segmenting Tuberculosis (TB)-consistent findings in frontal chest X-rays (CXRs). We determine an optimal uncertainty threshold based on several uncertainty-related metrics. This threshold is used to select and refer highly uncertain cases to an expert. Experimental results demonstrate that (i) the model trained with a modified Focal Tversky loss function delivered superior segmentation performance (mean average precision (mAP): 0.5710, 95% confidence interval (CI): (0.4021,0.7399)), (ii) the model with 30 MC forward passes during inference further improved and stabilized performance (mAP: 0.5721, 95% CI: (0.4032,0.7410), and (iii) an uncertainty threshold of 0.7 is observed to be optimal to refer highly uncertain cases.

Annotations of Lung Abnormalities in Shenzhen Chest X-ray Dataset for Computer-Aided Screening of Pulmonary Diseases.

Developments in deep learning techniques have led to significant advances in automated abnormality detection in radiological images and paved the way for their potential use in computer-aided diagnosis (CAD) systems. However, the development of CAD systems for pulmonary tuberculosis (TB) diagnosis is hampered by the lack of training data that is of good visual and diagnostic quality, of sufficient size, variety, and, where relevant, containing fine region annotations. This study presents a collection of annotations/segmentations of pulmonary radiological manifestations that are consistent with TB in the publicly available and widely used Shenzhen chest X-ray (CXR) dataset made available by the U.S. National Library of Medicine and obtained via a research collaboration with No. 3. People's Hospital Shenzhen, China. The goal of releasing these annotations is to advance the state-of-the-art for image segmentation methods toward improving the performance of fine-grained segmentation of TB-consistent findings in digital Chest X-ray images. The annotation collection comprises the following: 1) annotation files in JSON (JavaScript Object Notation) format that indicate locations and shapes of 19 lung pattern abnormalities for 336 TB patients; 2) mask files saved in PNG format for each abnormality per TB patient; 3) a CSV (comma-separated values) file that summarizes lung abnormality types and numbers per TB patient. To the best of our knowledge, this is the first collection of pixel-level annotations of TB-consistent findings in CXRs. Dataset: https://data.lhncbc.nlm.nih.gov/public/Tuberculosis-Chest-X-ray-Datasets/Shenzhen-Hospital-CXR-Set/Annotations/index.html.

Novel loss functions for ensemble-based medical image classification.

Medical images commonly exhibit multiple abnormalities. Predicting them requires multi-class classifiers whose training and desired reliable performance can be affected by a combination of factors, such as, dataset size, data source, distribution, and the loss function used to train deep neural networks. Currently, the cross-entropy loss remains the de-facto loss function for training deep learning classifiers. This loss function, however, asserts equal learning from all classes, leading to a bias toward the majority class. Although the choice of the loss function impacts model performance, to the best of our knowledge, we observed that no literature exists that performs a comprehensive analysis and selection of an appropriate loss function toward the classification task under study. In this work, we benchmark various state-of-the-art loss functions, critically analyze model performance, and propose improved loss functions for a multi-class classification task. We select a pediatric chest X-ray (CXR) dataset that includes images with no abnormality (normal), and those exhibiting manifestations consistent with bacterial and viral pneumonia. We construct prediction-level and model-level ensembles to improve classification performance. Our results show that compared to the individual models and the state-of-the-art literature, the weighted averaging of the predictions for top-3 and top-5 model-level ensembles delivered significantly superior classification performance (p < 0.05) in terms of MCC (0.9068, 95% confidence interval (0.8839, 0.9297)) metric. Finally, we performed localization studies to interpret model behavior and confirm that the individual models and ensembles learned task-specific features and highlighted disease-specific regions of interest. The code is available at https://github.com/sivaramakrishnan-rajaraman/multiloss_ensemble_models.

Improved Semantic Segmentation of Tuberculosis-Consistent Findings in Chest X-rays Using Augmented Training of Modality-Specific U-Net Models with Weak Localizations.

Deep learning (DL) has drawn tremendous attention for object localization and recognition in both natural and medical images. U-Net segmentation models have demonstrated superior performance compared to conventional hand-crafted feature-based methods. Medical image modality-specific DL models are better at transferring domain knowledge to a relevant target task than those pretrained on stock photography images. This character helps improve model adaptation, generalization, and class-specific region of interest (ROI) localization. In this study, we train chest X-ray (CXR) modality-specific U-Nets and other state-of-the-art U-Net models for semantic segmentation of tuberculosis (TB)-consistent findings. Automated segmentation of such manifestations could help radiologists reduce errors and supplement decision-making while improving patient care and productivity. Our approach uses the publicly available TBX11K CXR dataset with weak TB annotations, typically provided as bounding boxes, to train a set of U-Net models. Next, we improve the results by augmenting the training data with weak localization, postprocessed into an ROI mask, from a DL classifier trained to classify CXRs as showing normal lungs or suspected TB manifestations. Test data are individually derived from the TBX11K CXR training distribution and other cross-institutional collections, including the Shenzhen TB and Montgomery TB CXR datasets. We observe that our augmented training strategy helped the CXR modality-specific U-Net models achieve superior performance with test data derived from the TBX11K CXR training distribution and cross-institutional collections (p < 0.05). We believe that this is the first study to i) use CXR modality-specific U-Nets for semantic segmentation of TB-consistent ROIs and ii) evaluate the segmentation performance while augmenting the training data with weak TB-consistent localizations.

Modality-specific deep learning model ensembles toward improving TB detection in chest radiographs.

The proposed study evaluates the efficacy of knowledge transfer gained through an ensemble of modality-specific deep learning models toward improving the state-of-the-art in Tuberculosis (TB) detection. A custom convolutional neural network (CNN) and selected popular pretrained CNNs are trained to learn modality-specific features from large-scale publicly available chest x-ray (CXR) collections including (i) RSNA dataset (normal = 8851, abnormal = 17833), (ii) Pediatric pneumonia dataset (normal = 1583, abnormal = 4273), and (iii) Indiana dataset (normal = 1726, abnormal = 2378). The knowledge acquired through modality-specific learning is transferred and fine-tuned for TB detection on the publicly available Shenzhen CXR collection (normal = 326, abnormal =336). The predictions of the best performing models are combined using different ensemble methods to demonstrate improved performance over any individual constituent model in classifying TB-infected and normal CXRs. The models are evaluated through cross-validation (n = 5) at the patient-level with an aim to prevent overfitting, improve robustness and generalization. It is observed that a stacked ensemble of the top-3 retrained models demonstrates promising performance (accuracy: 0.941; 95% confidence interval (CI): [0.899, 0.985], area under the curve (AUC): 0.995; 95% CI: [0.945, 1.00]). One-way ANOVA analyses show there are no statistically significant differences in accuracy (P = .759) and AUC (P = .831) among the ensemble methods. Knowledge transferred through modality-specific learning of relevant features helped improve the classification. The ensemble model resulted in reduced prediction variance and sensitivity to training data fluctuations. Results from their combined use are superior to the state-of-the-art.

Detection and visualization of abnormality in chest radiographs using modality-specific convolutional neural network ensembles.

Convolutional neural networks (CNNs) trained on natural images are extremely successful in image classification and localization due to superior automated feature extraction capability. In extending their use to biomedical recognition tasks, it is important to note that visual features of medical images tend to be uniquely different than natural images. There are advantages offered through training these networks on large scale medical common modality image collections pertaining to the recognition task. Further, improved generalization in transferring knowledge across similar tasks is possible when the models are trained to learn modality-specific features and then suitably repurposed for the target task. In this study, we propose modality-specific ensemble learning toward improving abnormality detection in chest X-rays (CXRs). CNN models are trained on a large-scale CXR collection to learn modality-specific features and then repurposed for detecting and localizing abnormalities. Model predictions are combined using different ensemble strategies toward reducing prediction variance and sensitivity to the training data while improving overall performance and generalization. Class-selective relevance mapping (CRM) is used to visualize the learned behavior of the individual models and their ensembles. It localizes discriminative regions of interest (ROIs) showing abnormal regions and offers an improved explanation of model predictions. It was observed that the model ensembles demonstrate superior localization performance in terms of Intersection of Union (IoU) and mean Average Precision (mAP) metrics than any individual constituent model.

Analyzing inter-reader variability affecting deep ensemble learning for COVID-19 detection in chest radiographs.

Data-driven deep learning (DL) methods using convolutional neural networks (CNNs) demonstrate promising performance in natural image computer vision tasks. However, their use in medical computer vision tasks faces several limitations, viz., (i) adapting to visual characteristics that are unlike natural images; (ii) modeling random noise during training due to stochastic optimization and backpropagation-based learning strategy; (iii) challenges in explaining DL black-box behavior to support clinical decision-making; and (iv) inter-reader variability in the ground truth (GT) annotations affecting learning and evaluation. This study proposes a systematic approach to address these limitations through application to the pandemic-caused need for Coronavirus disease 2019 (COVID-19) detection using chest X-rays (CXRs). Specifically, our contribution highlights significant benefits obtained through (i) pretraining specific to CXRs in transferring and fine-tuning the learned knowledge toward improving COVID-19 detection performance; (ii) using ensembles of the fine-tuned models to further improve performance over individual constituent models; (iii) performing statistical analyses at various learning stages for validating results; (iv) interpreting learned individual and ensemble model behavior through class-selective relevance mapping (CRM)-based region of interest (ROI) localization; and, (v) analyzing inter-reader variability and ensemble localization performance using Simultaneous Truth and Performance Level Estimation (STAPLE) methods. We find that ensemble approaches markedly improved classification and localization performance, and that inter-reader variability and performance level assessment helps guide algorithm design and parameter optimization. To the best of our knowledge, this is the first study to construct ensembles, perform ensemble-based disease ROI localization, and analyze inter-reader variability and algorithm performance for COVID-19 detection in CXRs.

Iteratively Pruned Deep Learning Ensembles for COVID-19 Detection in Chest X-rays.

We demonstrate use of iteratively pruned deep learning model ensembles for detecting pulmonary manifestation of COVID-19 with chest X-rays. This disease is caused by the novel Severe Acute Respiratory Syndrome Coronavirus 2 (SARS-CoV-2) virus, also known as the novel Coronavirus (2019-nCoV). A custom convolutional neural network and a selection of ImageNet pretrained models are trained and evaluated at patient-level on publicly available CXR collections to learn modality-specific feature representations. The learned knowledge is transferred and fine-tuned to improve performance and generalization in the related task of classifying CXRs as normal, showing bacterial pneumonia, or COVID-19-viral abnormalities. The best performing models are iteratively pruned to reduce complexity and improve memory efficiency. The predictions of the best-performing pruned models are combined through different ensemble strategies to improve classification performance. Empirical evaluations demonstrate that the weighted average of the best-performing pruned models significantly improves performance resulting in an accuracy of 99.01% and area under the curve of 0.9972 in detecting COVID-19 findings on CXRs. The combined use of modality-specific knowledge transfer, iterative model pruning, and ensemble learning resulted in improved predictions. We expect that this model can be quickly adopted for COVID-19 screening using chest radiographs.

Performance evaluation of deep neural ensembles toward malaria parasite detection in thin-blood smear images.

BACKGROUND: Malaria is a life-threatening disease caused by Plasmodium parasites that infect the red blood cells (RBCs). Manual identification and counting of parasitized cells in microscopic thick/thin-film blood examination remains the common, but burdensome method for disease diagnosis. Its diagnostic accuracy is adversely impacted by inter/intra-observer variability, particularly in large-scale screening under resource-constrained settings. INTRODUCTION: State-of-the-art computer-aided diagnostic tools based on data-driven deep learning algorithms like convolutional neural network (CNN) has become the architecture of choice for image recognition tasks. However, CNNs suffer from high variance and may overfit due to their sensitivity to training data fluctuations. OBJECTIVE: The primary aim of this study is to reduce model variance, improve robustness and generalization through constructing model ensembles toward detecting parasitized cells in thin-blood smear images. METHODS: We evaluate the performance of custom and pretrained CNNs and construct an optimal model ensemble toward the challenge of classifying parasitized and normal cells in thin-blood smear images. Cross-validation studies are performed at the patient level to ensure preventing data leakage into the validation and reduce generalization errors. The models are evaluated in terms of the following performance metrics: (a) Accuracy; (b) Area under the receiver operating characteristic (ROC) curve (AUC); (c) Mean squared error (MSE); (d) Precision; (e) F-score; and (f) Matthews Correlation Coefficient (MCC). RESULTS: It is observed that the ensemble model constructed with VGG-19 and SqueezeNet outperformed the state-of-the-art in several performance metrics toward classifying the parasitized and uninfected cells to aid in improved disease screening. CONCLUSIONS: Ensemble learning reduces the model variance by optimally combining the predictions of multiple models and decreases the sensitivity to the specifics of training data and selection of training algorithms. The performance of the model ensemble simulates real-world conditions with reduced variance, overfitting and leads to improved generalization.

Pre-trained convolutional neural networks as feature extractors toward improved malaria parasite detection in thin blood smear images.

Malaria is a blood disease caused by the Plasmodium parasites transmitted through the bite of female Anopheles mosquito. Microscopists commonly examine thick and thin blood smears to diagnose disease and compute parasitemia. However, their accuracy depends on smear quality and expertise in classifying and counting parasitized and uninfected cells. Such an examination could be arduous for large-scale diagnoses resulting in poor quality. State-of-the-art image-analysis based computer-aided diagnosis (CADx) methods using machine learning (ML) techniques, applied to microscopic images of the smears using hand-engineered features demand expertise in analyzing morphological, textural, and positional variations of the region of interest (ROI). In contrast, Convolutional Neural Networks (CNN), a class of deep learning (DL) models promise highly scalable and superior results with end-to-end feature extraction and classification. Automated malaria screening using DL techniques could, therefore, serve as an effective diagnostic aid. In this study, we evaluate the performance of pre-trained CNN based DL models as feature extractors toward classifying parasitized and uninfected cells to aid in improved disease screening. We experimentally determine the optimal model layers for feature extraction from the underlying data. Statistical validation of the results demonstrates the use of pre-trained CNNs as a promising tool for feature extraction for this purpose.

Understanding the learned behavior of customized convolutional neural networks toward malaria parasite detection in thin blood smear images.

Convolutional neural networks (CNNs) have become the architecture of choice for visual recognition tasks. However, these models are perceived as black boxes since there is a lack of understanding of the learned behavior from the underlying task of interest. This lack of transparency is a serious drawback, particularly in applications involving medical screening and diagnosis since poorly understood model behavior could adversely impact subsequent clinical decision-making. Recently, researchers have begun working on this issue and several methods have been proposed to visualize and understand the behavior of these models. We highlight the advantages offered through visualizing and understanding the weights, saliencies, class activation maps, and region of interest localizations in customized CNNs applied to the challenge of classifying parasitized and uninfected cells to aid in malaria screening. We provide an explanation for the models' classification decisions. We characterize, evaluate, and statistically validate the performance of different customized CNNs keeping every training subject's data separate from the validation set.

Geographically distributed complementary content-based image retrieval systems for biomedical image informatics.

There is a significant increase in the use of medical images in clinical medicine, disease research, and education. While the literature lists several successful systems for content-based image retrieval and image management methods, they have been unable to make significant inroads in routine medical informatics. This can be attributed to the following: (i) the challenging nature of medical images, (ii) need for specialized methods specific to each image type and detail, (iii) lack of advances in image indexing methods, and (iv) lack of a uniform data and resource exchange framework between complementary systems. Most systems tend to focus on varying degrees of the first two items, making them very versatile in a small sampling of the variety of medical images but unable to share their strengths. This paper proposes to overcome these shortcomings by defining a data and resource exchange framework using open standards and software to develop geographically distributed toolkits. As proof-of-concept, we describe the coupling of two complementary geographically separated systems: the IRMA system at Aachen University of Technology in Germany, and the SPIRS system at the U. S. National Library of Medicine in the United States of America.

Image informatics at a national research center.

Image informatics at the Communications Engineering Branch of the Lister Hill National Center for Biomedical Communications (LHNCBC), an R&D division of the National Library of Medicine (NLM), includes document and biomedical images. In both domains, research into computer-assisted methods for information extraction, and the implementation of prototype systems incorporating such methods, is central to our mission. Current document image research focuses on extracting bibliographic data from scanned journal articles. Current biomedical imaging work focuses on content-based image retrieval (CBIR) and related problems in segmentation, indexing, and classifying collections of images of the spine and of the uterine cervix.

Biomedical image representation approach using visualness and spatial information in a concept feature space for interactive region-of-interest-based retrieval.

This article presents an approach to biomedical image retrieval by mapping image regions to local concepts where images are represented in a weighted entropy-based concept feature space. The term "concept" refers to perceptually distinguishable visual patches that are identified locally in image regions and can be mapped to a glossary of imaging terms. Further, the visual significance (e.g., visualness) of concepts is measured as the Shannon entropy of pixel values in image patches and is used to refine the feature vector. Moreover, the system can assist the user in interactively selecting a region-of-interest (ROI) and searching for similar image ROIs. Further, a spatial verification step is used as a postprocessing step to improve retrieval results based on location information. The hypothesis that such approaches would improve biomedical image retrieval is validated through experiments on two different data sets, which are collected from open access biomedical literature.

Towards the creation of a visual ontology of biomedical imaging entities.

Image content is frequently the target of biomedical information extraction systems. However, the meaning of this content cannot be easily understood without some associated text. In order to improve the integration of textual and visual information, we are developing a visual ontology for biomedical image retrieval. Our visual ontology maps the appearance of image regions to concepts in an existing textual ontology, thereby inheriting relationships among the visual entities. Such a resource creates a bridge between the visual characteristics of important image regions and their semantic interpretation. We automatically populate our visual ontology by pairing image regions with their associated descriptions. To demonstrate the usefulness of this resource, we have developed a classification method that automatically labels image regions with appropriate concepts based solely on their appearance. Our results for thoracic imaging terms show that our methods are promising first steps towards the creation of a biomedical visual ontology.