LEARN WHAT WE CONTRIBUTE TO MEDICINE


Lunit INSIGHT
for Chest Radiography
Nodule Detection

BACKGROUND

Proper detection of lung nodules, which includes lung cancer, is a challenging task when interpreting chest radiographs, with miss rates reported to be 20-30%.1,2 This is especially true for radiologists who need to read high volumes of images at limited amount of time, as well as for non-specialists who lack expertise in reading difficult cases, such as chest radiographs of small or hidden nodules. Missed lung cancer has serious clinical implications, with over 50% in reduction of 5-year survival rate when left undetected for around 1 year.3

PRODUCT DESCRIPTION

Developed using Lunit’s cutting-edge deep learning technology,4 Lunit INSIGHT CXR-Nodule accurately detects lung nodules in the form of diagnostic support tool. The AI solution generates (1) location information of detected lesions in the form of heatmaps and (2) abnormality scores reflecting the probability that the detected lesion is abnormal. The solution is indicated to be directly involved in the primary interpretation process of radiologists or clinicians as Second Reader.

PRIMARY VALUE PROPOSITION
  • Prevent nodules, especially small or hidden nodules, from being missed upon reading chest radiographs.
  • Help physicians make early diagnosis of lung cancer in chest radiographs.
  • Enable non-specialists to perform at specialist level in detecting lung nodules in chest radiographs.
TRAINING & VALIDATION
  • Trained with a large-scale (>70,000 cases), high-quality (clinically/CT-proven cases) training set.
  • Demonstrated to perform at a standalone accuracy of 97% in ROC AUC.5
  • Clinically validated to significantly improve the interpretive capabilities of clinicians and radiologists upto 20%.
  • Initial observer performance study published in Radiology.6
  • MFDS approved for clinical use in Korea (Computer-aided detection software; Approval No.18-574).
EXAMPLE CASES

CASE #1. A nodule, diagnosed as lung cancer, hidden behind the heart is properly detected, with an abnormality score of 44%. This case was missed by 8 out of 15 radiologists.

CASE #2. A nodule, diagnosed as lung cancer, in the right upper lung field is properly detected, with an abnormality score of 66%. This case was missed by 9 out of 9 radiologists.

CASE #3. A nodule, diagnosed as lung cancer, hidden behind the diaphragm is properly detected, with an abnormality score of 96%. This case was missed by 5 out of 9 radiologists.

JOURNALS & CONFERENCE ABSTRACTS
  • Nam JG, Park SG, et al. Development and Validation of Deep Learning-Based Automatic Detection Algorithm for Malignant Pulmonary Nodules on Chest Radiographs, Radiology 2018 (in press)
  • Development and Validation of a Deep Learning-Based Automatic Detection Algorithm for Active Pulmonary Tuberculosis on Chest Radiographs, RSNA 2018
  • Multi-Stage Deep Disassembling Networks for Generating Bone-Only and Tissue-Only Images from Chest Radiographs Performance Validation of a Deep Learning-Based Automatic, RSNA 2018
  • Detection Algorithm for Major Thoracic Abnormalities on Chest Radiographs, RSNA 2018
  • Deep Learning-Based Automatic Detection Algorithm for the Detection of Major Thoracic Abnormalities on Chest Radiographs, RSNA 2018
  • Automatic Detection of Malignant Pulmonary Nodules on Chest Radiographs Using a Deep Convolutional Neural Network: Detection Performance and Comparison with Human Experts, RSNA 2017
  • Deep Learning-based Automatic Detection Algorithm for the Detection of Malignant Pulmonary Nodules on Chest Radiographs, RSNA 2017
REFERENCES

1 Quekel LG, Kessels AG, Goei R, van Engelshoven JM. Miss rate of lung cancer on the chest radiograph in clinical practice. Chest 1999;115:720-4.

2 Forrest JV, Friedman PJ. Radiologic errors in patients with lung cancer. Wes J Med 1981;134:485.

3 Detterbeck FC, Gibson CJ. Turning gray: the natural history of lung cancer over time. J Thor Oncol 2008;3:781-92.

4 Lunit’s high-end deep learning technology has been demonstrated in various international competitions - won World #1 in MICCAI TUPAC 2016, and CAMELYON 2017; Recognized as one of the world's top 100 AI startups by CB Insights in 2017.

5 ROC AUC Area Under the Receiver Operating Characteristic Curve

6 Seoul National University Hospital, Observer performance study, 2017

7 Nam JG, Park SG, et al. Development and Validation of Deep Learning-Based Automatic Detection Algorithm for Malignant Pulmonary Nodules on Chest Radiographs. Radiology 2018; (in press).

Lunit INSIGHT
for Chest Radiography
Major Chest Abnormalities


BACKGROUND

Chest radiography is one of the most basic and fundamental diagnostic test used in medicine, accounting for 25% of the annual total numbers of diagnostic imaging procedures.1 It has been shown that radiologic information changed clinical practice in more than 60% of those who received chest radiography.2,3 Unfortunately, miss rates for proper interpretation of chest radiographs go as high as 30% even for experts,4,5 leading to increased mortality from treatable diseases.6 Moreover, interpretive performance of chest radiographs differ significantly between specialists and non-specialists, upto 30%.7-9 Among the various diseases detected or diagnosed through chest radiography, lung cancer (nodule/mass), tuberculosis, pneumonia (consolidation), and pneumothorax are among the most common and major diseases.

PRODUCT DESCRIPTION

Developed using Lunit’s cutting-edge deep learning technology,10 Lunit INSIGHT CXR-MCA accurately detects lung nodule/mass, consolidation, and pneumothorax in the form of diagnostic support tool. The AI solution generates (1) location information of detected lesions in the form of heatmaps and (2) abnormality scores reflecting the probability that the detected lesion is abnormal. The solution is indicated to be directly involved in the primary interpretation process of radiologists or clinicians as Second Reader.

PRIMARY VALUE PROPOSITION
  • Prevent difficult cases of major chest abnormalities from being missed upon reading chest radiographs.
  • Help physicians make early diagnosis of major chest abnormalities in chest radiographs.
  • Enable non-specialists to perform at specialist level in detecting major chest abnormalities in chest radiographs.
TRAINING & VALIDATION
  • Trained with a large-scale (>200,000 cases), high-quality (clinically/CT-proven cases) training set.
  • Demonstrated to perform at a standalone accuracy of 98-99% in ROC AUC.11
  • Clinically validated to significantly improve the interpretive capabilities of clinicians and radiologists upto 20%.
  • Currently in preparation for regulatory approval in various markets worldwide, including FDA, CE, MFDS.
EXAMPLE CASES

CASE #1. A small nodule, diagnosed as lung cancer, is properly detected in the right middle lung field, with an abnormality score of 94%.

CASE #2. Subtle focal consolidation, diagnosed as pneumonia, is properly detected in the right lower lung field, with an abnormality score of 81%.

CASE #3. Subtle pneumothorax is properly detected in the left apex, with an abnormality score of 57%.

CASE #4. Focal consolidation, diagnosed as tuberculosis, is properly detected in the right apex hidden behind the clavicle, with an abnormality score of 72%.

JOURNALS & CONFERENCE ABSTRACTS
  • Nam JG, Park SG, et al. Development and Validation of Deep Learning-Based Automatic Detection Algorithm for Malignant Pulmonary Nodules on Chest Radiographs, Radiology 2018 (in press)
  • Development and Validation of a Deep Learning-Based Automatic Detection Algorithm for Active Pulmonary Tuberculosis on Chest Radiographs, RSNA 2018
  • Multi-Stage Deep Disassembling Networks for Generating Bone-Only and Tissue-Only Images from Chest Radiographs Performance Validation of a Deep Learning-Based Automatic, RSNA 2018
  • Detection Algorithm for Major Thoracic Abnormalities on Chest Radiographs, RSNA 2018
  • Deep Learning-Based Automatic Detection Algorithm for the Detection of Major Thoracic Abnormalities on Chest Radiographs, RSNA 2018
  • Automatic Detection of Malignant Pulmonary Nodules on Chest Radiographs Using a Deep Convolutional Neural Network: Detection Performance and Comparison with Human Experts, RSNA 2017
  • Deep Learning-based Automatic Detection Algorithm for the Detection of Malignant Pulmonary Nodules on Chest Radiographs, RSNA 2017
REFERENCES

1 Radiation UNSCotEoA. Sources and effects of ionizing radiation: sources: United Nations Publications; 2000.

2 Geijer M, Ivarsson L, Göthlin JH. A retrospective analysis of the clinical impact of 939 chest radiographs using the medical records. Radiol Res Pract 2012;2012.

3 Speets AM, van der Graaf Y, Hoes AW, et al. Chest radiography in general practice: indications, diagnostic yield and consequences for patient management. Br J Gen Pract 2006;56:574-8.

4 Quekel LG, Kessels AG, Goei R, van Engelshoven JM. Miss rate of lung cancer on the chest radiograph in clinical practice. Chest 1999;115:720-4.

5 Forrest JV, Friedman PJ. Radiologic errors in patients with lung cancer. Wes J Med 1981;134:485.

6 Kesselman A, Soroosh G, Mollura DJ, et al. 2015 RAD-AID Conference on International Radiology for Developing Countries: the evolving global radiology landscape. J Am Coll Radiol 2016;13:1139-44.

7 Monnier-Cholley L, Carrat F, Cholley BP, Tubiana J-M, Arrivé L. Detection of lung cancer on radiographs: receiver operating characteristic analyses of radiologists’, pulmonologists’, and anesthesiologists’ performance. Radiology 2004;233:799-805.

8 Eng J, Mysko WK, Weller GE, et al. Interpretation of emergency department radiographs: a comparison of emergency medicine physicians with radiologists, residents with faculty, and film with digital display. Am J Roentgenol 2000;175:1233-8.

9 Potchen EJ, Cooper TG, Sierra AE, et al. Measuring performance in chest radiography. Radiology 2000;217:456-9.

10 Lunit’s high-end deep learning technology has been demonstrated in various international competitions - won World #1 in MICCAI TUPAC 2016, and CAMELYON 2017; Recognized as one of the world's top 100 AI startups by CB Insights in 2017.

11 ROC AUC Area Under the Receiver Operating Characteristic Curve

12 Seoul National University Hospital, Observer performance study, 201 7

Lunit INSIGHT
for Mammography


BACKGROUND

Screening mammography is the only single modality proven to improve breast cancer survival, and is the main test for breast cancer screening.1 However, screening breast cancer by mammography is far from ideal. False negative interpretations have been reported to range from 10 to 30%,2-4 and among the average 10% of subjects recalled for follow-up on initial screening mammogram, less than 5% are eventually diagnosed with cancer, representing a 95% false positive rate.5

PRODUCT DESCRIPTION

Developed using Lunit’s cutting-edge deep learning technology,6 Lunit INSIGHT MMG accurately detects lesions suspicious of breast cancer in the form of diagnostic support tool. The AI solution generates (1) location information of detected lesions in the form of heatmaps and (2) abnormality scores reflecting the probability that the detected lesion is malignant. The solution is indicated to be directly involved in the primary interpretation process of radiologists as Second Reader.

PRIMARY VALUE PROPOSITION
  • Assist both breast specialists and general radiologists (non-specialists) to increase cancer detection rate and decrease recall rate when interpreting screening mammograms.
  • Enable general radiologists to perform at specialist level.
TRAINING & VALIDATION
  • Trained with a large-scale (>200,000 total cases, >50,000 cancer cases), high-quality (biopsy-proven cases) training set.
  • Demonstrated to perform at a standalone accuracy of 96% in ROC AUC.7
  • Clinically validated to significantly improve the interpretive capabilities of radiologists upto 10%.8
  • Currently in preparation for regulatory approval in various markets worldwide, including FDA, CE, MFDS.
EXAMPLE CASES

CASE #1. A mass in the right breast, diagnosed as invasive ductal carcinoma, is shown to be properly detected with a malignancy score of 98%.

CASE #2. A lesion with microcalcifications in the right breast, diagnosed as ductal carcinoma in situ, is shown to be properly detected with a malignancy score of 58%.

JOURNALS & CONFERENCE ABSTRACTS
  • Kim EK et al. Applying Data-driven Imaging Biomarker in Mammography for Breast Cancer Screening: Preliminary Study, Sci Rep. 2018 Feb 9;8(1):2762.
  • Kim EK et al. Data-Driven Imaging Biomarker for Breast Cancer Screening in Mammography - Reader Study, RSNA 2018
  • Advanced Data-Driven Imaging Biomarker for Breast Cancer Screening in Mammography, RSNA 2017
REFERENCES

1 Myers ER, Moorman P, Gierisch JM, et al. Benefits and harms of breast cancer screening: a systematic review. JAMA 2015;314:1615-34.

2 Thurfjell EL, Lernevall KA, Taube A. Benefit of independent double reading in a population-based mammography screening program. Radiology 1994;191:241-4.

3 Yankaskas BC, Klabunde CN, Ancelle-Park R, et al. International comparison of performance measures for screening mammography: can it be done? J Med Screen 2004;11:187-93.

4 Ciatto S, Ambrogetti D, Risso G, et al. The role of arbitration of discordant reports at double reading of screening mammograms. J Med Screen 2005;12:125-7.

5 http://breastscreening.cancer.gov.

6 Lunit’s high-end deep learning technology has been demonstrated in various international competitions - won World #1 in MICCAI TUPAC 2016, and CAMELYON 2017; Recognized as one of the world's top 100 AI startups by CB Insights in 2017.

7 ROC AUC Area Under the Receiver Operating Characteristic Curve

8 Severance Hospital, Observer performance study, 2017

BACK TO TOP