Bao Do

baodo @ stanford edu 

 

 

 

A picture containing indoor, cup, table, glasses

Description automatically generated  A picture containing film, indoor, sitting, table

Description automatically generated  Air 11:Users:Bao:Desktop:Screen Shot 2017-08-14 at 10.32.40 AM.jpg  A picture containing indoor, mirror, photo, white

Description automatically generated Pooter:Users:baodo:Desktop:Screen Shot 2017-11-08 at 3.31.19 PM.jpeg

 

 

APPLICATIONS:

 

“A.I.” to automatically write essays

 

radTF, a Radiology report search engine @ Stanford (1, 2).

 

clariPACS.com. Cloud PACS for research, teaching, embedding DICOM in webpages, email sharing, peer learning. Univ of Calgary Stroke, Stanford MSK & Neuroradiology cases.

 

UNITY. Radiology operations and automatic protocolling for the VA.

 

Omega. RIS for imaging operation and deep learning in the VA, eg pre-dictate x-ray in workflow. Try neural networks here.

 

Bone tumor Bayesian network link.

 

OCAD MSK link.

 

Musculoskeletal MRI atlas link.

 

 

CLINICAL, EDUCATION, & RESEARCH PROJECTS:

 

Keypoint Annotator

Tool for fast annotation of keypoints for training neural networks to learn angles and distances btwn points in an image (see below, “Automatic Diagnosis of Knee Patella Malalignment…”). Will use to map skeletal landmarks of the human body to automate quantitative reporting of bone x-rays.

 

Aryan Kaul, Bao Do.

 

Graphical user interface, text, application

Description automatically generated

Report Miner

Custom search engine indexing 15+ years of VistA radiology reports at VA Palo Alto for education and clinical work.

 

Charles Fang, Andrew Wu, Bao Do.

Original Report miner article

 

 

A close-up of a human brain

Description automatically generated with low confidence

 

Expert level detection of slipped capital femoral epiphysis using artificial intelligence

Andrew Campion, Audrey Ha, Bao Do, Charles Fang, Kevin Shea, Michael Fadell. AI for detection of SCFE in 3 grades, using only the AP view (no frog leg). In small test set, AI 99% accurate (103/104 cases).

 

 

A picture containing person, child, young, person

Description automatically generated

 

Automatic Extraction of Skeletal Maturity from Whole Body Pediatric Scoliosis X-rays.

Skeletal maturity assessment plays an important role in the management of pediatric orthopedic conditions such as scoliosis, slipped capital femoral epiphysis (SCFE), and pectus. The most common methods to estimate bone age are the use of hand, shoulder, and/or pelvis x-ray. Can whole body x-rays be helpful for extracting skeletal maturity estimates (ie, Oxford stage).

 

Audrey Ha, John Vorhies, Andrew Campion, Charles Fang, Michael Fadell II, Steve Dou, Safwan Halabi, David Larson, Emily Wang, YongJin Lee, Joanna Langner, Japsimran Kaur, Bao Do. Short paper publication. IEEE BIBM Conference 2020.

 

A picture containing wearing, standing, close, person

Description automatically generated

 

Automatic Diagnosis of Knee Patella Malalignment on X-ray Using Artificial Intelligence. Measurement of the Install-Salvati, modified Install-Salvati, and Caton-Deschamps Index on knee xray exams.

 

Aryan Kaul, Jason Pai, Charles Fang, Ed Boas, Kathryn Stevens, Jason Saleh, Vananh Nguyen, Constance Chu, Jamie Schroder, Michelle Nguyen, Joshua Reicher, Woon Teck Yap, Amelie Lutz, Bao Do. AMIA 2020 Annual Symposium.

 

 

A picture containing table, photo, sitting, small

Description automatically generated

 

Automatic Detection of Aortic Aneurysm on CT Exams Using Deep Convolutional Neural Networks.

Audrey Ha, Charles Fang, Saif Baig, Bao Do.

AMIA 2020 Annual Symposium, November, 2020.

 

 

A picture containing photo, man, cat

Description automatically generated

 

Pediatric ChestNet: Neural network for detetion of pneumonia on pediatric chest x-rays

Applying the public data from Kermany et al.’s excellent work, “Identifying Medical Diagnoses and Treatable Diseases by Image-Based Deep Learning”, which trained AI to diagnosis of pediatric pneumonia using chest x-ray images. In our limited test of 60 unknowns (20 normal, 20 viral, 20 bacterial pneumonia), AI missed 1 pna, 4 false positives; F1 score 0.94, sensitivity 97.5% and accuracy 91.6%. Using Win 2016 2GB RAM CPU only, Pediatric ChestNet reads @ ~8s/film. Try it: Download image1 or image2 & try the AI

 

A picture containing indoor, mirror, photo, white

Description automatically generated

 

Automatic detetion of thigh sarcoma on whole body PET/CT

Limited test 14 (9 tumors + 5 normals), AI 92.8% accurate, F1 score 0.94. Lays foundation for AI based detection, interpretatation, and reporting of PET and CT/MRI exams.

A picture containing clothing, underwear, grey, gray

Description automatically generated

 

Precision diagnosis of bone tumors on radiography using deep learning and bayesian network demo

AI for generating ddx of bone tumors on xray using computer vision and a bayesian probabilistic network

 

A picture containing film, indoor, sitting, table

Description automatically generated

Automatic interpretation and dication of scoliosis exams on radiography using machine learning

"Each year, >400,000 office, 130,000 hospital, and 17,000 emergency room visits are made by children with scoliosis. 29,000 surgeries are performed yearly in adolescents, with an avg hospital stay cost of $92,000. Scoliosis patients are up to 40% more likely to have suicidal thoughts." -ACR Data Science Institute. Try it: Download image or PNG/JPG fr your PACS & try the AI

 

 

Radiology mangement software for the Veteran’s Health Administration. Joshua Reicher, Payam Massaband, B. Do link

 

Macintosh HD:Users:MacPro:Desktop:Screen Shot 2018-11-22 at 12.54.08 AM.jpg

 

Deep learning of wrist bone xray exams for trauma.

Machine Learning-Aided Diagnosis Enhances Human Detection of Perilunate Dislocations. 75th ASSH Annual Meeting, San Francisco. October 1-3, 2020.

Pooter:Users:baodo:Desktop:1.jpg Pooter:Users:baodo:Desktop:segond study:training_images:segonds:scvmc-export--317445581.jpg

 

“Expert” level perception of ACL trauma on knee xray exams using sparse data. A.I. automatic diagnosis of a subtle x-ray finding (Segond fracture, less than 0.04% of entire area of an image) using contextual interpretation, like a radiologist. Machine was 100% accurate (F-score 1.0) in our limited 20 test samples. Science poster presentation RSNA 2018.

 

A picture containing holding, woman, hand, wearing

Description automatically generated

 

Detection and characterization of peri-articular calcification using deep learning

Potentially for automatic quantification and calcinosis staging.

 

Automatic chondroid bone tumor detection using deep learning

Can machines identify tumors such as chondroid mass ? This is a simple patch based CNN that was slow and not particularly sensitive or specific.

 

 

Machine learning decision trees in radiology

Experiment using A.I. to automatically build expert decision trees from radiology observations to partition tumor classes. 100% accuracy in limited testing but broad application (classify 10+ tumor dxs, 15+ observations per dx) was plagued w/ over-fitting due to small sample size at this time.

 

 

IVC filter machine learning

Teach computer to recognize IVC filters on x-ray images [ article ].

 

 

 

Pooter:Users:baodo:Desktop:unnamed.png

 

 

Chondroid machine

Adapted Stanford Bone Bayes engine to discriminate benign (enchondroma) VS malignant (chondrosarcoma) chondroid bone tumors. Preliminary performance exceeded >85% of radiologists. Do, BH, Beaulieu CF, Human VS Machine: Distinguishing Enchondroma from Chondrosarcoma with a Bayesian Network, Society of Skeletal Radiology meeting, 3/2018. Rec’d SSR “Man Vs Machine” Award, 2018.

 

Air 11:Users:Bao:Desktop:Screen Shot 2017-08-14 at 10.36.23 AM.jpg

 

DeepBone

Deep convolutional neural net to identify bone joints of the human body. Foundation for PACS comparison protocols, automated decision systems such bone tumor diagnosis, information retrieval, QA, etc. Dr. Vossler rec’d a Student Travel Award for Young Investigators at the RSNA 2017 Scientific Meeting.

 

Air 11:Users:Bao:Desktop:Screen Shot 2017-08-14 at 10.32.40 AM.jpg

 

Stanford Bone Bayes

A “learning” Bayesian network that models clinical and radiographic inputs to compute diagnosis, differential diagnosis, and probabilities. link publication

 

 

Air 11:Users:Bao:Desktop:Screen Shot 2017-08-14 at 10.48.21 AM.jpg

clariPACS

Co-founded secure cloud PACS that can be customized for clinical care, education, research, media reports. More than 100,000 users have now viewed > 10 million images on clariPACS. Highlights: Univ of Calgary Stroke findings and Stanford Neuroradiology

 

 

Pooter:Users:baodo:Desktop:Screen Shot 2017-11-08 at 1.30.55 PM.jpeg

OCAD

Community of radiologists promoting the exchange of high level musculoskeletal imaging expertise. Searchable gallery of common and rare musculoskeletal disorders for decision support, teaching. Founded by Philip Tirman MD (2003). Served > 1 million images since 2010.

 

 

Pooter:Users:baodo:Desktop:Screen Shot 2017-11-08 at 2.52.00 PM.jpeg

 

Stanford Atlas of Common MSK Measurements

RadLex-enabled atlas of common MSK measurements. Designed to leverage RadLex ID interoperability for automation systems. RSNA Scientific Meeting, 2010.

 

 

Pooter:Users:baodo:Desktop:Screen Shot 2017-11-08 at 1.36.57 PM.jpeg

ACGME Analytics

To eliminate the labor intensive task of counting Radiology residency workloads at VA centers by mining VistA to generate an ACGME compliant summary of workload by CPTs. The Accreditation Council for Graduate Medical Education (ACGME) requires bi-annual case log statistical reports by Residency programs for all medical specialties. The Residency Review Committee (RRC) for Radiology recognizes 90 CPT codes in 11 categories: chest radiographs, CT body, CT angiography, image guided biopsy and drain, mammography, MRI body, MRI brain, MRI knee, PET/CT, ultrasound pelvis, and MRI spine.

 

Pooter:Users:baodo:Desktop:Screen Shot 2017-11-08 at 1.46.21 PM.jpeg

Hedge Detector

Automatic extraction of uncertainty and recommendation concepts from unstructured radiology reports. Applications include for QA, follow-up, peer review, disease surveillance. Rec’d RSNA Research Trainee Prize for Scientific Paper, RSNA Scientific Meeting, 2009.

 

Pooter:Users:baodo:Desktop:Screen Shot 2017-11-08 at 1.47.52 PM.jpeg

VA Radiology Analytics

Mine and summarize operations metrics for VA Hospitals using VistA radiology data. Common metrics to help identify bottlenecks in radiology clinical operations at Veteran hospitals.

 

Pooter:Users:baodo:Desktop:Screen Shot 2017-11-08 at 1.49.17 PM.jpeg

 

Stanford MSK MRI Atlas

Free MRI atlas of musculoskeletal radiology.

Has served > 600,000 pages & > 20 million+ images to users from 100+ countries since 2011. RSNA Scientific Meeting, 2010.

 

Pooter:Users:baodo:Desktop:Screen Shot 2017-11-08 at 1.54.29 PM.jpeg

 

RadTF

Radiology report search engine, originally indexed nearly 1,000,000 cases in Stanford radiology. Shared for free to numerous academic institutions, hospitals, and individuals around the world, including deployment at Hospital de Pediatria J. P. Garrahan, a children’s hospital in Buenos Aires, Argentina (Darío Filippo MD), where it enables search of over 250,000 radiology, pathology, and surgery notes. Interesting behaviors in medical imaging search. RSNA Scientific Meeting, 2009.

 

 

 

Pooter:Users:baodo:Desktop:Screen Shot 2017-11-08 at 1.56.46 PM.jpeg

 

iKeyNote

Audience response system w/ MCQ style interactive polling for Stanford radiology lectures.

 

 

Pooter:Users:baodo:Desktop:Screen Shot 2017-11-08 at 2.04.30 PM.jpeg

 

NLP to structure MSK MRI knee reporting

Auto structured knee MRI reports from free dictation in real time.

RSNA Scientific Meeting, 2010.

 

Pooter:Users:baodo:Desktop:Screen Shot 2017-11-08 at 2.10.04 PM.jpeg

 

NLP to detect missing concepts in MSK tumor reporting

May augment self learning systems via extraction of ground truth correlation of RIS/PATHOLOGY. RSNA Scientific Meeting, 2012.

 

Pooter:Users:baodo:Desktop:Screen Shot 2017-11-08 at 2.11.57 PM.jpeg

NLP to automate quality assurance using RIS reports - Fractures

Fully automatic extraction of accuracy statistics from unstructured radiology report data using NLP to compare knee x-ray reads against the follow-up MRI/CT reports as ground truth. QA, education, teaching, peer review. RSNA Scientific Meeting, 2011.

 

 

Pooter:Users:baodo:Desktop:Screen Shot 2017-11-08 at 2.16.23 PM.jpeg

 

NLP to automate quality assurance using RIS reports - Osteoporosis

Automatically extract accuracy metrics for osteoporosis from radiology reports in the RIS, using NLP to compare against the follow-up DEXA reports, respectively, as ground truth. RSNA Scientific Meeting, 2011.

 

Pooter:Users:baodo:Desktop:Screen Shot 2017-11-08 at 2.14.25 PM.jpeg

 

NLP to derive central line usage from unstructured ICU x-ray reports

Using NLP to build an ICU patient consensus and central line indwelling estimate from unstructured chest x-ray reports.

 

 

Pooter:Users:baodo:Desktop:Screen Shot 2017-11-08 at 2.17.49 PM.jpeg

 

Feedback NLP of fractures in unstructured reports of ED studies

Real-time retrieval of fracture classifications and clinical pearls extracted from free dictation for decision support. Rec’d RSNA Research Trainee Prize for Scientific Paper, RSNA Scientific Meeting, 2007.

 

Pooter:Users:baodo:Desktop:Screen Shot 2017-11-08 at 3.31.25 PM.jpegPooter:Users:baodo:Desktop:Screen Shot 2017-11-08 at 3.31.19 PM.jpeg

 

Characterizing radiology search patterns using web analytics

System to analyze how radiologists review studies using heat maps of mouse movement and magnification patterns, deployed completely via the web. Note how experienced, attending radiologists (LEFT) require fewer mouse movements and magnification compared to trainee radiology residents (RIGHT). Deployed in online simulator of call cases for residents. RSNA Scientific Meeting, 2008.

 

 

AWARDS:      

                                                                                                             

2005. RSNA Research Trainee Prize for Scientific Paper. "Metabolic Profile of the Spinal Cord by Whole Body [18F]-2-deoxy-2-fluoro-d-glucose (18F-FDG) Positron Emission Tomography (PET) / Computed Tomography (CT) Imaging".

2007. Outstanding Radiology Resident Teacher, University of Iowa.

2007. RSNA Research Trainee Prize for Scientific Paper. "Feedback Natural Language Processing of Fractures in Unstructured Reports of Emergency Department Studies".

2009. RSNA Research Trainee Prize for Scientific Paper. “A Natural Language Processor to Detect Uncertainty and Recommendations in Radiology Reports”.

2010. RSNA Roentgen Resident Research Award, Stanford University Medical Center.

2011. Fellow Teaching Award, Stanford Univeristy Medical Center.

2012. Clinical Educator of the Year, Stanford Hospital and Clinics, CA.

2018. VA Palo Alto Healthcare Director’s Commendation for patient care/UNITY, Oct 5 2018, w/ Joshua Reicher and Payam Massaband.

2018: Stanford-Philips Grant. Precision Diagnosis of Bone Tumors with Advanced Semantic Modeling and Radiomics Analysis. Co-PI: Christopher Beaulieu, MD PhD.

 

 

SELECTED PUBLICATIONS:          

 

Wu, A. S., Do, B. H., Kim, J., & Rubin, D. L. (2009). Evaluation of Negation and Uncertainty Detection and its Impact on Precision and Recall in Search. Journal of Digital Imaging, 24(2), 234–242. http://doi.org/10.1007/s10278-009-9250-4

 

Do, B. H., Wu, A., Biswal, S., Kamaya, A., & Rubin, D. L. (2010). Informatics in Radiology: RADTF: A Semantic Search–enabled, Natural Language Processor–generated Radiology Teaching File. RadioGraphics, 30(7), 2039–2048. http://doi.org/10.1148/rg.307105083

 

Do, B. H., Wu, A. S., Maley, J., & Biswal, S. (2012). Automatic Retrieval of Bone Fracture Knowledge Using Natural Language Processing. Journal of Digital Imaging, 26(4), 709–713. http://doi.org/10.1007/s10278-012-9531-1

 

De-Arteaga, M., Eggel, I., Do, B., Rubin, D., Kahn, C. E., Jr., & Müller, H. (2015). Comparing image search behaviour in the ARRS GoldMiner search engine and a clinical PACS/RIS. Journal of Biomedical Informatics, 56, 57–64. http://doi.org/10.1016/j.jbi.2015.04.013

 

Do, BH, Langlotz, C, Beaulieu, CF. Bone Tumor Diagnosis Using a Naïve Bayesian Model of Demographic and Radiographic Features. Journal of Digitial Imaging. 2017 Jul 27. doi: 10.1007/s10278-017-0001-7.

 

Banerjee, I., Kurtz, C., Edward Devorah, A., Do, B., Rubin, D. L., Beaulieu, C. F. Relevance Feedback for Enhancing Content Based Image Retrieval and Automatic Prediction of Semantic Image Features: Application to Bone Tumor Radiographs. Journal of biomedical informatics. 2018.

 

Ni JC, Shpanskaya K, Han M, Lee EH, Do BH, Kuo WT, Yeom KW, Wang DS. Deep-Learning for Automated Classification of Inferior Vena Cava Filter Types on Radiographs. J Vasc Interv Radiol. 2019 Sep 18. pii: S1051-0443(19)30536-6. doi: 10.1016/j.jvir.2019.05.026.

 

Enamandram, S. Sandhu, E. Do, BH, Reicher, JJ., Beaulieu, CF. Artificial Intelligence and Machine Learning Applications in Musculoskeletal Imaging. Advances in Clinical Radiology. 28 May 2020. https://doi.org/10.1016/j.yacr.2020.05.005

 

Callen, AL. Dupont, SM., Price, A., Laguna, B., McCoy, D., Do, BH, Talbott, J. Kohli, M., Narvid, J. Between Always and Never: Evaluating Uncertainty in Radiology Reports Using Natural Language Processing. Journal of Digital Imaging. August 2020. https://doi.org/10.1007/s10278-020-00379-1.

 

Audrey Ha, John Vorhies, Andrew Campion, Charles Fang, Michael Fadell II, Steve Dou, Safwan Halabi, David Larson, EmilyAutomatic Extraction of Skeletal Maturity from Whole Body Pediatric Scoliosis X-rays Using Regional Proposal and Compound Scaling Convolutional Neural Networks. Short paperpublication. IEEE International Conference on Bioinformatics and Biomedicine (BIBM), 12/2020.