This will count as one of your downloads.
You will have access to both the presentation and article (if available).
Methods: 4 phantoms were investigated: CDMAM, L1, CIRS BR3D and Modular DBT Phantom (two different inserts). The phantoms were imaged on recent DBT models: Fujifilm Amulet Innovality (ST mode), GE HC Senographe Pristina, Hologic 3Dimensions, IMS Giotto Class and Siemens Mammomat Revelation. Images were acquired at automatic exposure control (AEC) level, half AEC and twice AEC. SM was calculated. The CDMAM and L1 phantom were read by human readers via a 4-Alternative Forced Choice method and thresholds were established. CIRS BR3D and Modular DBT Phantom were analysed by counting visible lesions.
Results: The scores obtained from the phantoms had the same tendencies among systems. The phantoms highlight many specific characteristics of the SM algorithms such as tuning contrast enhancement to a range of sizes. The phantoms confirm, as in 2D and DBT, an impact of dose on detectability of microcalcification-like inserts but not on masses. None of the phantoms evaluate the SM for different glandular tissue or thickness distributions.
Conclusion: For all phantoms, SM found a number of lesion-like targets and an impact of dose as expected. Whether these phantom readings are representative for quality in SM in real practice is not yet proven. More elaborated sensitivity studies should be done prior to the use of the phantoms in routine QC. Ultimately, accurate assessment of SM may have to be done via virtual trials.
We utilize PCA to sample input breast cases, then by using weighted sums along the different eigenvectors or "eigenbreasts," a number of new cases can be generated. While breasts can vary in structure and form, we used a series of compressed breasts derived from human subject breast CT volumes to create the eigenbreasts. We used an initial set of thirty-five phantoms from a new CT patient population with 155x155x155 μm3 voxel size. The training set and synthetized phantoms were evaluated by power law exponent β and changes in volumetric breast density as a result of the PCA process.
The synthetic phantoms were found to have similar β and fibroglandular density distributions to the training dataset. Individual synthetic phantoms appeared to capture glandular features present in the training phantoms but had visually different texture features. This work shows that earlier work on the eigenbreast technique can be extended to newer datasets with higher resolution and produce synthetic phantoms that retain the quantitative properties of training data.
The main contribution of our work can be summarized as follows. 1) The proposed one-class learning requires only data from one class, i.e., the negative data; 2) The patch-based learning makes the proposed method scalable to images of different sizes and helps avoid the large scale problem for medical images; 3) The training of the proposed deep convolutional neural network (DCNN) based auto-encoder is fast and stable.
Method: In this study, we collected digital mammography magnification views for 140 patients with DCIS at biopsy, 35 of which were subsequently upstaged to invasive cancer. We utilized a deep CNN model that was pre-trained on two natural image data sets (ImageNet and DTD) and one mammographic data set (INbreast) as the feature extractor, hypothesizing that these data sets are increasingly more similar to our target task and will lead to better representations of deep features to describe DCIS lesions. Through a statistical pooling strategy, three sets of deep features were extracted using the CNNs at different levels of convolutional layers from the lesion areas. A logistic regression classifier was then trained to predict which tumors contain occult invasive disease. The generalization performance was assessed and compared using repeated random sub-sampling validation and receiver operating characteristic (ROC) curve analysis.
Result: The best performance of deep features was from CNN model pre-trained on INbreast, and the proposed classifier using this set of deep features was able to achieve a median classification performance of ROC-AUC equal to 0.75, which is significantly better (p<=0.05) than the performance of deep features extracted using ImageNet data set (ROCAUC = 0.68).
Conclusion: Transfer learning is helpful for learning a better representation of deep features, and improves the prediction of occult invasive disease in DCIS.
Images were acquired on a Hologic Selenia Dimensions system with a uniform and anthropomorphic phantom. A contrast detail insert of small, low-contrast disks was created using an inkjet printer with iodine-doped ink and inserted in the phantoms. The disks varied in diameter from 210 to 630 μm, and in contrast from 1.1% contrast to 2.2% in regular increments. Human and model observers performed a 4-alternative forced choice experiment. The models were a non-prewhitening matched filter with eye model (NPWE) and a channelized Hotelling observer with either Gabor channels (Gabor-CHO) or Laguerre-Gauss channels (LG-CHO).
With the given phantoms, reader scores were higher in FFDM and DBT than SM. The structure in the phantom background had a bigger impact on outcome for DBT than for FFDM or SM. All three model observers showed good correlation with humans in the uniform background, with ρ between 0.89 and 0.93. However, in the structured background, only the CHOs had high correlation, with ρ=0.92 for Gabor-CHO, 0.90 for LG-CHO, and 0.77 for NPWE.
Because results of any analysis can depend on the phantom structure, conclusions of modality performance may need to be taken in the context of an appropriate model observer and a realistic phantom.
Estimating breast density with dual energy mammography: a simple model based on calibration phantoms
This will count as one of your downloads.
You will have access to both the presentation and article (if available).
View contact details
No SPIE Account? Create one
