Search

In the development of a computer-aided detection (CAD) system a large database of training samples is of major importance. However digital breast tomosynthesis (DBT) is a relatively new modality and no large database of cases is available yet. To overcome this limitation we are developing a CAD system for mass detection in DBT that can be trained with regular 2D mammograms, for which large datasets are available. We trained our system with a very large database of screen-film mammograms (SFM). Our approach does not use projection images, but only reconstructed volumes, because it is expected that manufacturers of tomosynthesis systems will only store the reconstructed volumes. In this study we developed a method that converts reconstructed volumes into a series of SFM-like slices and combinations of slices, called slabs. By combining slices into slabs, more information of a whole mass, which usually spans several slices, is used and its appearance becomes more similar to a 2D mammogram. In this study we investigate the effect of using slabs of different sizes on the performance of our CAD system. For validation we use a dataset of 63 tomosynthesis cases (245 volumes) consisting of 42 normal cases (163 volumes) and 21 abnormal cases (82 volumes) with a total of 47 malignant masses and architectural distortions. The volumes are acquired with a tomosynthesis system from Sectra and are reconstructed into 0.3 cm thick slices. Results show that performance of our CAD system increases significantly when slices are combined into larger slabs. Best performance is obtained when a slab thickness of 1.5 cm (5 slices) is used, which is significantly higher than using slabs of a single slice, two slices and all slices. [ABSTRACT FROM AUTHOR]

Ideally, the outcome of any CAD performance assessment should predict how well the system would work if used clinically. In principle, if the selection process draws cases that are ˵representative″ of the general patient population, the study design will be unbiased. In this study we explored the effect of stratified sampling on stand-alone and radiologists΄ performance using data from an observer study. Although our database was relatively small, 50 cancer cases, no meaningful difference in performance was measured among different stratified sampling schemes or against the whole dataset nor was there any difference in the variance in the measured performance metrics. These results cast doubts on the usefulness of requiring stratified sampling, whose added cost does not seem to be justifiable without empirical evidence. We believe that it is more important to specify how cases should be collected than try to define the range and frequency of the characteristics of patients and cancers to be included the dataset, which we suspect to be prone to actually produce unrealistic samples. [ABSTRACT FROM AUTHOR]

Breast cancer is one of the major causes of death among women all over the world. Presently, mammographic analysis is the most used method for early detection of abnormalities. This paper presents a computational methodology to help the specialist with this task. In the first step, the K-Means clustering algorithm and the Template Matching technique are used to detect suspicious regions. Next, the texture of each region is described using the Simpson΄s Diversity Index, which is used in Ecology to measure the biodiversity of an ecosystem. Finally, the information of texture is used by SVM to classify the suspicious regions into two classes: masses and non-masses. The tests demonstrate that the methodology has 79.12% of accuracy, 77.27% of sensitivity, and 79.66% of specificity. [ABSTRACT FROM AUTHOR]

We are developing a computer-aided detection (CAD) system for masses on digital breast tomosynthesis mammograms (DBT). A data set of DBT from patients was collected with informed consent using a GE prototype DBT system. The system has an Rh/Rh x-ray source and a CsI/a:Si detector, and acquires 21 projection view (PV) images over a 60° arc in 3° increments. The total dose of the 21 PVs is set to be about 1.5 times of the dose of a single screen-film mammogram. In this preliminary study, the DBTs from 50 patients are reconstructed with a simultaneous algebraic reconstruction technique (SART) in two conditions: one using all 21 PVs in 3° increments and the other 11 PVs in 6° increments. The latter therefore uses a subset of the PVs and about half the dose of the former. We compared the mass detection accuracy in these two sets of DBTs using the same CAD system. The system first identifies locations of mass candidates in the DBT volume using 3D gradient field analysis. The mass candidates are then segmented by 3D region growing. Morphological, gray level, and texture features are extracted from the segmented objects. A linear discriminant analysis (LDA) classifier with stepwise feature selection is designed by leave-one-out resampling. The output of the LDA is a mass likelihood score to which a decision threshold is applied and a free-response receiver operating characteristic (FROC) curve is generated. For the DBT set reconstructed with 21 PVs, the system achieved an 80% sensitivity at an average false positive (FP) rate of 1.52 per DBT volume. For the DBT set reconstructed with 11 PVs, the FP rate was 2.38 per breast at the same sensitivity. The difference in the FROC curves for the two conditions was statistically significant (p=0.03) by alternative FROC analysis. When the subset of 14 malignant cases was analyzed, the average FP rates were 0.87 and 1.11 per DBT volume at a sensitivity of 80% for the 21-PV and 11-PV conditions, respectively. The difference in the FROC curves for this smaller subset did not achieve statistical significance. [ABSTRACT FROM AUTHOR]

A focal asymmetric density (FAD), chiefly identified by comparing left breast and right breast mammograms, is a common finding related to breast cancers detected in mammograms. In our database, 18.4% of cases with any findings were identified as FADs. It is often difficult for radiologists to distinguish FADs from specific conditions of glandular tissue, and FADs are the second most frequently missed or misinterpreted reason for cases with interval cancers. Therefore, the purpose of this study was to develop a computerized method for detection of FADs in mammograms in order to assist earlier diagnoses of FADs by radiologists. Our computerized method for detection of FAD is based on geometric and texture analyses, and consisted of eight steps. Sixty-four images (32 FAD images and 32 normal images) were used for training and testing an artificial neural network (ANN) which distinguishes FADs from normal tissues. The performance of our computerized method was evaluated by a round-robin test. 84.4% of images with FAD (27/32) were correctly detected by the computerized method at 0.56 false positives per image. Our preliminary results show that the automated computerized scheme for detecting FAD on mammograms has the potential to assist radiologists in detecting FAD. [ABSTRACT FROM AUTHOR]