Kurt Rossmann Laboratories

or Radiologic Image Research


ROC Software
Contact Us




Thoracic Computed Tomography (CT)
1. Automated Detection of Lung Nodules in CT Scans

We developed a fully automated computerized method for the detection of lung nodules in helical CT scans of the thorax.  This method was based on 2D and 3D analyses of the image data acquired during diagnostic CT scans.  Lung segmentation proceeded on a section-by-section basis to construct a segmented lung volume within which further analysis was performed.  Multiple gray-level thresholds were applied to the segmented lung volume to create a series of thresholded lung volumes.  An 18-point connectivity scheme was used to identify contiguous 3D structures within each thresholded lung volume, and those structures that satisfy a volume criterion were selected as initial lung nodule candidates.  Morphological and gray-level features were computed for each nodule candidate.  After a rule-based approach was applied to reduce the number of nodule candidates that corresponds to non-nodules, the features of remaining candidates were merged through linear discriminant analysis.  The automated method was applied to a database of 43 diagnostic thoracic CT scans.  ROC analysis was used to evaluate the ability of the classifier to differentiate nodule candidates that corresponded to actual nodules from false-positive candidates.  The Az value for this categorization task was 0.90 during leave-one-out-by-case evaluation.  The automated method yielded an overall nodule detection sensitivity of 70% with an average of 1.5 false-positive detections per section.  A corresponding nodule detection sensitivity of 89% with an average of 1.3 false-positive detections per section was achieved with a subset of 20 cases that contained only one or two nodules per case.  (409, 587, 629, 630, 632, 654, 695, 745, 746, 795)


2.  Analysis of Missed Lung Cancers in Low-Dose Helical CT Screening

From May 1996 to March 1999, 17,892 examinations with low-dose helical CT were performed in Nagano, Japan.  Each CT scan was read initially by one of five general radiologists to identify suspicious abnormalities due to lung cancer or non-cancerous lesions.  Another chest radiologist then reviewed the CT images of these suspicious cases to make a final diagnosis.  Eighty-three primary lung cancers were histologically confirmed.  Thirty-two missed lung cancers were identified in 39 scans, which included detection errors in 23 scans and interpretation errors in 16 scans.  The correlation of clinical characteristics, CT features, and pathologic findings of these missed lung cancers was examined.  All missed cancers were peripherally located.  In the detection error group, 55% (11/20) were non-smoking females, and 93% (18/20) of cancers were well-differentiated adenocarcinomas.  The mean size of cancers missed due to detection error (9.8 mm) was less than that due to interpretation error (15.9 mm).  The percentage of faint nodules with ground-glass opacity (GGO) in the detection error group (91%=21/23) was greater than that in the interpretation error group (38%=6/16).  In the detection error group, 91% (21/23) of the lesions were considered subtle or were very subtle, and 83% (19/23) were adjacent to, overlapped with, or similar in appearance to normal structures such as pulmonary vessels.  Ninety-four percent (15/16) of interpretation errors were due to findings similar to benign lesions or to confusing findings caused by other abnormalities such as tuberculosis and inflammatory lesions.  In conclusion, missed lung cancers in this series of low-dose CT screening examination were generally very subtle and appeared as small faint nodules, overlapping normal structures, or opacities in a complex background caused by other diseases.  (822)


3. Performance of Automated CT Nodule Detection on Missed Cancers

We evaluated the performance of the computerized method for the detection of lung nodules in CT scans in the identification of lung cancers that were missed during visual interpretation.  A database of 38 low-dose CT scans with 50 lung nodules was obtained from a lung cancer screening program.  Thirty-eight of the nodules represented biopsy-confirmed lung cancers that were not reported during initial clinical interpretation.  Our computer method uses gray-level-thresholding techniques to identify three-dimensionally contiguous structures within the lungs.  At a specific operating point on the FROC curve, the method achieved a sensitivity of 80.0% (40 out of 50 nodules) with an average of 1.0 false-positive detections per section.  Missed cancers were detected by the computerized method with a sensitivity of 84.2% (32 out of 38 missed cancers) and a false-positive rate of 1.0 per section.  With an automated lung nodule detection method, a large fraction of missed cancers in a database of low-dose CT scans was detected correctly.  (795)



Detection errorSensitivity78%
FPs per section1.6
Number of nodules23
Interpretation error Sensitivity93%
FPs per section1.6
Number of nodules 15


Results: Missed cancers (leave-one-out by patient analysis)

4. Low-Dose CT Screening for Lung Cancer in a General Population: Characteristics of Cancer in Non-Smokers versus Smokers

To report the detection rate for lung cancers in CT screening in Japanese adults, and to analyze differences in the appearance of the cancers in non-smokers versus smokers. Subjects consisted of 7,847 Japanese adults, who received low-dose CT screening at least once in a three-year period. The detection rate of lung cancers and the correlation of imaging, clinical, and pathologic findings of cancers in non-smokers versus smokers were examined. The detection rate for lung cancer was 1.1% for both non-smokers (45/4,251) and smokers (39/3,596). The prevalence of well-differentiated adenocarcinomas was greater in non-smokers (88%=22/25) than in smokers (29%=4/14) (P<0.001). The prevalence and incidence of pathologic stage IA disease were greater in non-smokers than in smokers (92%=22/24 versus 58%=7/12, and 100%=19/19 vs. 70%=14/20) (both P<0.05). The mean size of the tumors in the non-smokers (12.4 mm) was smaller than that in smokers (18.2 mm) (P<0.001). The percentage of cancers categorized as pure or mixed ground-glass opacity (GGO) (86%=38/44) on CT was greater in non-smokers than in smokers (46%=16/35) (P<0.001). Most of the lung cancers in non-smokers were slowly growing adenocarcinomas appearing as faint GGOs on CT, whereas rapidly growing cancers appearing as solid nodules were more commonly seen in smokers. (871)


5. Reduction of False Positives by Use of Massive Training Artificial Neural Network (MTANN)

We developed a massive training artificial neural network (MTANN) for reduction of false positives in computerized detection of lung nodules in low-dose CT images.  The MTANN consisted of a modified multilayer ANN, which was capable of operating on image data directly.  The MTANN was trained by use of input images together with the teacher images containing the distribution for the “likelihood of being a nodule.”  To achieve high performance, the MTANN was trained by using a large number of sub-regions extracted from an input image.  The output image was obtained by scanning an input image with the MTANN, i.e., the MTANN acted like the convolution kernel of a filter.  The distinction between a nodule and a non-nodule was made by use of a score which was defined from the output image of the trained MTANN.  The MTANN was applied to the task of reducing the number of false positives reported when our current computerized scheme for lung nodule detection was applied to a database of 50 nodules, including 38 “missed” cancers.  The MTANN was trained with ten typical nodules and ten typical non-nodules and then applied to the remaining 40 nodules and 1068 non-nodules (false positives) in a validation test.  The results indicated that 66% (706/1068) of the false positives were removed without a reduction in the number of true positives, i.e., a sensitivity of 100% (40 out of 40 nodules) with 0.35 false positives per section.  In addition, the Multi-MTANN consisting of ten MTANNs were trained with ten typical nodules and ten non-nodules representing each of ten different non-nodule types (100 training non-nodules overall).  The Multi-MTANN was then applied to the remaining (i.e., non-training) cases of 40 nodules and 978 non-nodules in a validation test.  The results indicated that 93% (905/978) of non-nodules were removed without a reduction in the number of true positives.  By using the Multi-MTANN, the false-positive rate of our current scheme could be improved from 1.02 to 0.08 false positives per section, while potentially maintaining a current high sensitivity.  (890, 891)


6. Computer-Aided Diagnosis for Detection of Missed Peripheral Lung Cancers on CT: ROC and LROC Analysis

To evaluate whether a computer-aided diagnostic (CAD) scheme can assist radiologists in detecting missed peripheral lung cancers on CT. Seventeen low-dose CT scans with a missed peripheral lung cancer and 10 CT scans without a cancer were employed in an observer study. Fourteen radiologists were divided into two groups based on use of different image display formats: Six radiologists (Group 1) reviewed CT images on a multi-format display, and eight radiologists (Group 2) on a “stacked” cine mode. The radiologists, first without and then with the CAD scheme, indicated their confidence level regarding the presence (or absence) of a cancer, and also the most likely position of a lesion on each CT scan. Receiver operating characteristic (ROC) curves, without and with localization, were obtained for evaluating of the observers’ performance. With the aid of the CAD scheme, the average Az value for the ROC curve improved from 0.763 to 0.854 for all radiologists (P = 0.002); from 0.757 to 0.862 for Group 1 (P = 0.04) and from 0.768 to 0.848 for Group 2 (P = 0.01). The sensitivity for cancer detection improved from 52% to 68% (P < 0.0001), from 49% to 70% for Group1 (P = 0.02) and from 54% to 67% for Group 2 (P = 0.006). The localization ROC (LROC) curve also improved. Lung cancers missed at low-dose CT screening were very difficult to detect, even in an observer study. However, CAD can improve radiologists’ performance in detecting these subtle cancers. (966)

ROC curves for detection of lung nodules in low-dose CT

7. Likelihood of Malignancy of Pulmonary Nodules on Low-Dose Helical CT

We developed an automated computerized scheme for determination of the likelihood of malignancy of pulmonary nodules on low-dose helical CT.  Our database consisted of 76 primary lung cancers and 413 benign nodules, which were obtained from a lung cancer screening on 7,847 screenees with a low-dose helical CT in Nagano, Japan.  With this automated computerized scheme, the location of a nodule is first indicated by a radiologist.  The nodule outline was determined automatically, and forty-three image features were determined from quantitative analysis of the outline, texture, and gray-level histogram on the segmented nodule region.  A linear discriminant analysis (LDA) was employed to distinguish benign from malignant nodules using 43 features and two clinical parameters (age and sex).  Many different combinations of 45 features were examined as input to the LDA.  Results indicated that the Az value obtained by the computerized scheme in distinguishing benign from malignant nodules was 0.840, which was greater than the Az value of 0.625 obtained by the radiologist alone.  The automated computerized scheme for determination of the likelihood of nodule malignancy would be useful in assisting radiologists in their task of distinguishing between benign and malignant solitary pulmonary nodules on low-dose helical CT.  (845)


8. CAD Scheme for Distinction between Benign and Malignant Nodules in Thoracic Low-Dose CT by Use of MTANN

Low-dose helical CT (LDCT) is being applied as a modality for lung cancer screening. It may be difficult, however, for radiologists to distinguish malignant from benign nodules in LDCT. Our purpose in this study was to develop a computer-aided diagnostic (CAD) scheme for distinction between benign and malignant nodules in LDCT scans by use of a massive training artificial neural network (MTANN). The MTANN is a trainable, highly nonlinear filter based on an artificial neural network. To distinguish malignant nodules from six different types of benign nodules, we developed a multiple MTANN (Multi-MTANN) consisting of six expert MTANNs that are arranged in parallel. Each of the MTANNs was trained by use of input CT images and teacher images containing the distribution for the "likelihood of being a malignant nodule," i.e., the teacher image for a malignant nodule contains a two-dimensional Gaussian distribution and that for a benign nodule contains zero. Each MTANN was trained independently with ten typical malignant nodules and ten benign nodules from each of the six types. The outputs of the six MTANNs were combined by use of an integration ANN such that the six types of benign nodules could be distinguished from malignant nodules. After training of the integration ANN, our scheme provided a value related to the “likelihood of malignancy” of a nodule, i.e., a higher value indicates a malignant nodule, and a lower value indicates a benign nodule. Our database consisted of 76 primary lung cancers in 73 patients and 413 benign nodules in 342 patients, which were obtained from a lung cancer screening program on 7,847 screenees with LDCT for three years in Nagano, Japan. The performance of our scheme for distinction between benign and malignant nodules was evaluated by use of ROC analysis. Our scheme achieved an Az value of 0.882 in a round-robin test. Our scheme correctly identified 100% (76/76) of malignant nodules as malignant, whereas 48% (200/413) of benign nodules were identified correctly as benign. Therefore, our scheme would be useful in assisting radiologists in the diagnosis of lung nodules in LDCT. (954)

ROC curves indicating the performance of a multi-MTANN incorporating an integration ANN

9. Improvement in Radiologists’ Performance for Differentiating Small Benign from Malignant Lung Nodules on High-Resolution CT by Using Computer-Estimated Likelihood of Malignancy

The purpose of our study was to evaluate whether a computer-aided diagnosis (CAD) scheme can assist radiologists in distinguishing small benign from malignant nodules on high-resolution CT (HRCT). We developed an automated computerized scheme for determination of the likelihood of malignancy of lung nodules on multiple HRCT slices, which was obtained from various objective features of the nodules by use of linear discriminant analysis. The data set used in this observer study consisted of 28 primary lung cancers (6-20 mm) and 28 benign nodules. Cancer cases included nodules with pure ground glass opacity (GGO), mixed GGO, and solid opacity. Benign nodules were selected by matching of their size and pattern to the cancers. Consecutive region of interest (ROI) images for each nodule on HRCT were displayed for interpretation in stacked mode on a CRT monitor. The images were presented to sixteen radiologists, first without and then with the computer output, to indicate their confidence level regarding the malignancy of a nodule, and the performance was evaluated by receiver operating characteristic (ROC) analysis. The Az value of the CAD scheme alone was 0.831 for distinguishing benign from malignant nodules. The average Az value for radiologists was improved from 0.785 to 0.853 by a statistically significant level (P=0.016) with the aid of the CAD scheme. The radiologists’ performance with the CAD scheme was greater than that of the CAD scheme alone (P<0.05), and also that of radiologists alone. CAD has the potential to improve diagnostic accuracy in distinguishing small benign nodules from malignant ones on HRCT. (926)

ROC curves for distinction between benign and malignant nodules on high-resolution CT

10. Subtraction CT Technique for Detection of Lung Nodules

We developed a new subtraction CT technique to enhance isolated structures such as subtle pulmonary nodules and to assist radiologists for the detection of lung cancer on low-dose CT images.  In this scheme, the mask image for a given section was obtained by use of 3D morphological filtering technique applied to adjacent sections immediately above and below the section in question.  An iterative image warping technique was then applied on the mask image for registration of normal structures such as vessels and heart.  The subtraction CT images were obtained by subtracting the warped mask images from the corresponding CT images.  Results indicated that the subtraction CT technique could remove the majority of normal background structures extended over multiple slices such as ribs, vessels, and heart.  Thus, isolated structures were clearly enhanced as dark shadows on subtraction CT images.  Subtraction CT images would be useful for the detection of lung cancers on CT images, and thus could assist radiologists in the early detection of lung cancer. (899)


11. Effect of Temporal Subtraction Images on Detection of Lung Cancer

Temporal subtraction is an effective image processing technique which can enhance selectively areas of interval changes by using the previous image as a subtraction mask.  The purpose of this study was to evaluate the effect of temporal subtraction on low-dose CT on the detection of early primary lung cancer by an observer study.  We selected 14 cases with primary lung cancer and 16 normal cases for this study from our database of low-dose CT images, which were obtained from a CT screening program for lung cancer.  Temporal subtraction technique included a global matching technique, in which the previous image was adjusted its position in order to correct for the difference between the previous and the current image, and an iterative image warping technique, in which the previous image was nonlinearly warped for the best match with the current image.  Eleven radiologists, including five attendings and six residents, participated in this study.  Results indicated that diagnostic accuracy was significantly improved with the use of temporal subtraction images.  The mean Az values of the observers without and with temporal subtraction were 0.88 and 0.93, respectively (P<0.05).  Temporal subtraction images were especially useful when the nodule was present near the pulmonary hilum, where radiologists tended to overlook the nodule. (899)


12. Intelligent CAD Scheme for Distinction between Benign and Malignant Nodules

An intelligent CAD scheme for distinction of pulmonary nodules provides not only the likelihood of malignancy but also sets of benign and malignant images similar to an unknown nodule.  First, the location of a nodule was identified by a chest radiologist, and the nodule outline was determined automatically by use of a dynamic programming technique.  Forty-three features were determined from the original image and edge-gradient image based on the segmented nodule region.  A LDA and an ANN were employed, respectively, to determine the likelihood of malignancy and to select the similar images for an unknown nodule.  Our database included 76 primary cancers and 413 benign nodules obtained from a lung cancer screening program.  Sixteen radiologists participated in our observer study with 20 benign nodules and 20 malignant nodules randomly selected from the large database.  The Az value obtained with the computed likelihood of malignancy was 0.83.  By use of the intelligent CAD scheme, all radiologists improved their diagnostic performance, with a significant increase of the average Az value from 0.72 to 0.80 (P<0.0001).  The intelligent CAD scheme can significantly improve the radiologists’ performance in distinction between benign and malignant nodules in low-dose CT scans. (873)

13. Selective Enhancement Filters for Nodules, Vessels, and Airway Walls in Two- and Three-Dimensional CT Scans

Computer-aided diagnostic (CAD) schemes have been developed to assist radiologists in the early detection of lung cancer in radiographs and computed tomography (CT) images. In order to improve sensitivity for nodule detection, many researchers have employed a filter as a preprocessing step for enhancement of nodules. However, these filters enhance not only nodules, but also other anatomic structures such as ribs, blood vessels, and airway walls. Therefore, nodules are often detected together with a large number of false positives caused by these normal anatomic structures. In this study, we developed three selective enhancement filters for dot, line, and plane which can simultaneously enhance objects of a specific shape (for example, dot-like nodules) and suppress objects of other shapes (for example, line-like vessels). Therefore, as preprocessing steps, these filters would be useful for improving the sensitivity of nodule detection and for reducing the number of false positives. We applied our enhancement filters to synthesized images to demonstrate that they can selectively enhance a specific shape and suppress other shapes. We also applied our enhancement filters to real two-dimensional (2-D) and three-dimensional (3-D) CT images to show their effectiveness in the enhancement of specific objects in real medical images. The three enhancement filters developed in this study would be useful in the computerized detection of cancer in 2-D and 3-D medical images. (874)

Original Image and  Nodule-enhanced image with a subtle nodule

Home | Up | History | ROC Software | Publications | Contact Us

This site was last updated 11/24/04