Document text
Principal Investigator: Sameer Antani
Organization: NATIONAL LIBRARY OF MEDICINE
Fiscal Year: 2024
Award: $1,370,493
Funding agency: National Library of Medicine
ML/AI driven automated computer-aided diagnostic (CADx) tools are designed to detect, localize, classify and grade disease in medical images to augment human-expert decision-making, add efficiencies, and improve overall performance. However, reliability of the predictions depends significantly on several training data characteristics which include quality, imbalance between cases and controls, volume, variety, and the truth standard (human assessment or histopathology). The prediction performance and reliability also depends on deep learning networks used, architecture, use of multimodality, and if synthetic data was used in training and how it might be incorporated into the learning and prediction process. Toward this, we focused our research on various medical image analysis tasks such as quality assessment, image enhancement, region of interest detection and segmentation, image classification and prediction interpretation. Several advances were made to address these topics through applications to diseases of interest. We continued our research on ensemble learning techniques and multimodal (text + image) learning which promise to provide benefits from combining the predictions from different models and result in improved generalizability and overall accuracy. We also added novel generative AI research to synthesize high quality images. Advances in these areas were embedded in various disease-detection driven ML/AI research described below.
It is commonly believed that deep learning methods are data hungry and more data would mean better outcomes. Toward this, we studied if blindly increasing complex and multidimensional data for training would result in improved performance. However, we noted that with medical imaging data there exists signifcant semantic redundancy, which is the presence of similar or repetitive information, that can occur due to the presence of multiple images that have highly similar presentations for the disease of interest. We proposed an entropy-based sample scoring approach to identify and remove semantically redundant training data and demonstrate using the publicly available NIH chest X-ray dataset that the model trained on the resulting informative subset of training data significantly outperforms the model trained on the full training set, during both internal and external testing. Our findings emphasized the importance of information-oriented training sample selection as opposed to the conventional practice of using all available training data.
Infectious diseases detected on chest x-rays: Automated segmentation of tuberculosis (TB)-consistent lesions in CXRs using DL methods can help reduce radiologist effort and supplement clinical decision-making. This is particularly significant in children living with or without HIV. Toward this, in collaboration with NIAID, we solicited annotations (indicators or disease) on over 1500 frontal and lateral pediatric chest x-ray images from three radiologists with expertise in pediatric TB.
We participated in a challenge to detect COVID-19 induced pulmonary edema on chest x-rays. Self supervised contrastive learning techniques were used to successfully detect the disease. Our results were ranked in the top-10 submitted runs, and top-5 in the fully automatic solutions. We also studied the effects of model initialization for deep learning. While much literature exists on non-medical images, the impacts on medical images, particularly chest X-rays (CXRs) are less understood. Addressing this gap, our study explores three deep model initialization techniques: Cold-start, Warm-start, and Shrink and Perturb start, focusing on adult and pediatric populations. We specifically focused on scenarios with periodically arriving data for training, thereby embracing the real-world scenarios of ongoing data influx and the need for model updates. We evaluated these models for generalizability against external adult and pediatric CXR datasets. We also proposed novel ensemble methods. Our evaluations indicated that models initialized with ImageNet-pretrained weights demonstrated superior generalizability over randomly initialized counterparts, contradicting some findings for non-medical images.
One of the common problems with medical imaging AI is its inability to generalize to data from different sources. We wanted to investigate this issue to gain insights into its characteristics toward determining which parameters are controllable toward minimize its impact. Toward this, we analyzed domain shift in lung region detection on five chest X-ray datasets, collected from different public sources. We compared the characteristics of these datasets from three aspects: information obtained from metadata or an image header, image appearance, and features extracted from a pretrained model. We proposed a new feature visualization method to provide explanations for the applied object detection network on the obtained quantitative results. We also examined chest X-ray modality-specific initialization, catastrophic forgetting, and model repeatability.
Cardiovascular disease: Automated echocardiography (echo) analysis is benefited through use of machine learning for tasks such as image quality assessment, view classification, cardiac region segmentation, and quantification of diagnostic indices. Our novel and efficient DL-based real-time system for echo analysis and quantification has acquired international patent protection. Efforts are underway to make the code available to interested parties. Next steps in this work include plans to expand the work to those afflicted with sickle cell disease.
Cervical cancer: Continuing research with NCI, we supported their efforts in inducing robustness, reliability, and generalizability across devices and different populations for women living with and without HIV. in AI through training and evaluation. Our efforts in this project are now sunset.
Oral cavity malignant lesion analysis: Oral cavity cancer is a common cancer that can result in significant impairments, and there is high mortality for the advanced stage. The final diagnosis is confirmed through histopathology, however high variability is observed among human experts in determining if a subject needs biopsy and identifying the correct biopsy location. Further, the disease can occur in different parts of the oral cavity. Toward developing an ML-based method that can help address these problems and reduce downstream classification errors, we automatically identify, with high accuracy, different anatomical sites in the oral cavity on the images that are verified using class activation maps obtained from both correct and incorrect predictions.
This year we continued efforts on data quality control, and deep learning to predict cancerous lesions in buccal tissue. A NIDCR fellow (dentist with oral oncological expertise) annotated lesions which were used for training a deep learning model which predicted 60% false positive rate at 90% true positive rate. If used in the field this would mean a reduction in 40% biopsies which are currently at 100% (ie. every patient with suspected precancerous lesions is subjected to biopsy verification). Additional research is ongoing.
Kaposi Sarcoma on Dark-skinned population: We developed methods to identify and grade dark skin and separate it from irrelevant background in the image. We also evaluated several methods to automatically locate potential KS lesions and then classify as them as being KS or not. Initial results in the study demonstrated improvement over human experts for detecting low-grade and high grade lesions. The results were comparable to human ambivalence for mid-grade lesions. Additional research is ongoing.
Terms: <0-11 years old><21+ years old><AI algorithm><AI system><AIDS Virus><Acquired Immune Deficiency Syndrome Virus><Acquired Immunodeficiency Syndrome Virus><Address><Adult><Adult Human><Affect><Anatomic Sites><Anatomic structures><Anatomy><Appearance><Architecture><Area><Artificial Intelligence><Biomedical Research><Biopsy><Body Tissues><Buccal Cavity><Buccal Cavity Head and Neck><COVID detection><COVID-19 detection><COVID19 detection><CXR><Cancerous><Cancers><Cardiac><Cardiovascular Diseases><Cavitas Oris><Cervical Cancer><Cervix Cancer><Characteristics><Child><Child Youth><Childhood><Children (0-21)><Classification><Clinical><Code><Coding System><Collaborations><Communicable Diseases><Complex><Computer Reasoning><Conventional X-Ray><Darkness><Data><Data Set><Decision Making><Dentists><Detection><Devices><Diagnosis><Diagnostic><Disease><Disorder><Echocardiogram><Echocardiography><Echography><Echotomography><Engineering / Architecture><Entropy><Evaluation><General Population><General Public><Goals><HIV><Hb SS disease><HbSS disease><Hemoglobin S Disease><Hemoglobin sickle cell disease><Hemoglobin sickle cell disorder><Histopathology><Human><Human Immunodeficiency Viruses><Image><Image Analyses><Image Analysis><Image Enhancement><Imaging Device><Imaging Instrument><Imaging Tool><Impairment><Infectious Disease Pathway><Infectious Diseases><Infectious Disorder><International><Kaposi Sarcoma><Kaposi's Sarcoma><LAV-HTLV-III><Lateral><Learning><Legal patent><Lesion><Literature><Location><Low-resource area><Low-resource community><Low-resource environment><Low-resource region><Low-resource setting><Lung><Lung Respiratory System><Lymphadenopathy-Associated Virus><M tuberculosis infection><M. tb infection><M. tuberculosis infection><M.tb infection><M.tuberculosis infection><MTB infection><Machine Intelligence><Machine Learning><Malignant><Malignant - descriptor><Malignant Cervical Neoplasm><Malignant Cervical Tumor><Malignant Neoplasm of the Cervix><Malignant Neoplasms><Malignant Oral Cavity Neoplasm><Malignant Oral Cavity Tumor><Malignant Oral Neoplasm><Malignant Tumor><Malignant Tumor of the Cervix><Malignant Tumor of the Cervix Uteri><Malignant Uterine Cervix Neoplasm><Malignant Uterine Cervix Tumor><Malignant neoplasm of cervix uteri><Maps><Medical Imaging><Medical Ultrasound><Metadata><Methods><Modality><Modeling><Modern Man><Mouth><Mouth Cancer><Multiple Hemorrhagic Sarcoma><Mycobacterium tuberculosis (MTB) infection><Mycobacterium tuberculosis infection><NIAID><NIDCR><NIDR><NIH><National Institute of Allergy and Infectious Disease><National Institute of Dental Research><National Institute of Dental and Craniofacial Research><National Institutes of Health><Oncology><Oncology Cancer><Oral><Oral Cancer><Oral cavity><Outcome><Patents><Patients><Performance><Population><Population Heterogeneity><Process><Property><Pulmonary Edema><Quality Control><Real-Time Systems><Research><Resource-constrained area><Resource-constrained community><Resource-constrained environment><Resource-constrained region><Resource-constrained setting><Resource-limited area><Resource-limited community><Resource-limited environment><Resource-limited region><Resource-limited setting><Resource-poor area><Resource-poor community><Resource-poor environment><Resource-poor region><Resource-poor setting><Role><Running><SARS-CoV-2 detection><Sampling><Semantics><Sickle Cell Anemia><Skin><Source><Systematics><TB infection><Techniques><Testing><Text><Thoracic Radiography><Tissues><Training><Transthoracic Echocardiography><Tuberculosis><Ultrasonic Imaging><Ultrasonogram><Ultrasonography><Ultrasound Diagnosis><Ultrasound Medical Imaging><Ultrasound Test><United States National Institutes of Health><Update><Uterine Cervix Cancer><Variant><Variation><Virus-HIV><Visualization><Weight><Woman><Work><X-Ray Imaging><X-Ray Medical Imaging><Xray imaging><Xray medical imaging><adulthood><artificial intelligence algorithm><auto-segmentation><automated segmentation><automatic segmentation><autosegmentation><biomedical imaging><cardiovascular disorder><chest X ray><chest Xray><chest radiography><clinical decision-making><co-morbid><co-morbidity><comorbidity><computer-aided diagnostic><computer-assisted diagnostics><conventional Xray><coronavirus detection><coronavirus disease 2019 detection><coronavirus disease detection><data quality><data to train><dataset to train><deep learning><deep learning based model><deep learning method><deep learning model><deep learning strategy><design><designing><detect COVID><detect COVID-19><detect COVID19><detect SARS-CoV-2><detect coronavirus><detect coronavirus disease><detect severe acute respiratory syndrome coronavirus 2><diagnostic tool><diagnostic ultrasound><disseminated TB><disseminated tuberculosis><diverse populations><forgetting><generative AI><generative artificial intelligence><global health><health application><heart sonography><heterogeneous population><high dimensional data><image evaluation><image interpretation><image processing><imaging><improved><indexing><infection due to Mycobacterium tuberculosis><innovate><innovation><innovative><insight><interest><kids><learning network><lung edema><lung radiography><machine based learning><machine learning based method><machine learning method><machine learning methodologies><malignancy><malignant mouth neoplasm><malignant mouth tumor><meta data><model generalizability><mortality><multi-modality><multidimensional data><multidimensional datasets><multimodality><neoplasm/cancer><novel><oral cavity cancer><pediatric><population diversity><portability><pre-trained model><precancer><precancerous><premalignant><pulmonary><radiographic chest image><radiographic lung image><radiologist><realtime systems><self supervised><self supervised learning><self supervision><severe acute respiratory syndrome coronavirus 2 detection><sickle cell disease><sickle cell disorder><sickle disease><sicklemia><social role><sonogram><sonography><sound measurement><synthetic data><task analysis><thoracic radiogram><thorax radiography><training data><treatment planning><tuberculosis infection><tuberculous spondyloarthropathy><ultrasound imaging><ultrasound scanning><weights><youngster>