AI-enhanced Mammography With Digital Breast Tomosynthesis for Breast Cancer Detection: Clinical Value and Comparison With Human Performance.

Purpose To compare two deep learning-based commercially available artificial intelligence (AI) systems for mammography with digital breast tomosynthesis (DBT) and benchmark them against the performance of radiologists. Materials and Methods This retrospective study included consecutive asymptomatic patients who underwent mammography with DBT (2019-2020). Two AI systems (Transpara 1.7.0 and ProFound AI 3.0) were used to evaluate the DBT examinations. The systems were compared using receiver operating characteristic (ROC) analysis to calculate the area under the ROC curve (AUC) for detecting malignancy overall and within subgroups based on mammographic breast density. Breast Imaging Reporting and Data System results obtained from standard-of-care human double-reading were compared against AI results with use of the DeLong test. Results Of 419 female patients (median age, 60 years [IQR, 52-70 years]) included, 58 had histologically proven breast cancer. The AUC was 0.86 (95% CI: 0.85, 0.91), 0.93 (95% CI: 0.90, 0.95), and 0.98 (95% CI: 0.96, 0.99) for Transpara, ProFound AI, and human double-reading, respectively. For Transpara, a rule-out criterion of score 7 or lower yielded 100% (95% CI: 94.2, 100.0) sensitivity and 60.9% (95% CI: 55.7, 66.0) specificity. The rule-in criterion of higher than score 9 yielded 96.6% sensitivity (95% CI: 88.1, 99.6) and 78.1% specificity (95% CI: 73.8, 82.5). For ProFound AI, a rule-out criterion of lower than score 51 yielded 100% sensitivity (95% CI: 93.8, 100) and 67.0% specificity (95% CI: 62.2, 72.1). The rule-in criterion of higher than score 69 yielded 93.1% (95% CI: 83.3, 98.1) sensitivity and 82.0% (95% CI: 77.9, 86.1) specificity. Conclusion Both AI systems showed high performance in breast cancer detection but lower performance compared with human double-reading. Keywords: Mammography, Breast, Oncology, Artificial Intelligence, Deep Learning, Digital Breast Tomosynthesis © RSNA, 2024.

Technical note: Realization and uncertainty analysis for an adjustable 3D structured breast phantom in digital breast tomosynthesis.

BACKGROUND: Projection imaging phantoms are often optimized for 2-dimensional image characteristics in homogeneous backgrounds. Therefore, evaluation of image quality in tomosynthesis (DBT) lacks accepted and established phantoms. PURPOSE: We describe a 3D breast phantom with a structured, variable background. The phantom is an adaptable and advanced version of the L1 phantom by Cockmartin et al. Phantom design and its use for quality assurance measurements for DBT devices are described. Four phantoms were compared to assess the objectivity. METHODS: The container size was increased to a diameter of 24 cm and a total height of 53.5 mm. Spiculated masses were replaced by five additional non-spiculated masses for higher granularity in threshold diameter resolution. These patterns are adjustable to the imaging device. The masses were printed in one session with a base layer using two-component 3D printing. New materials compared to the L1 phantom improved the attenuation difference between the lesion models and the background. Four phantoms were built and intra-human observer, inter-human observer and inter-phantom variations were determined. The latter assess the reproducibility of the phantom production. Coefficients of variance (V) were calculated for all three variations. RESULTS: The difference of the attenuation coefficients between the lesion models and the background was 0.20 cm-1 (with W/Al at 32 kV, equivalent to 19-20 keV effective energy) compared to 0.21 cm-1 for 50/50 glandular/adipose breast tissue and cancerous lesions. PMMA equivalent thickness of the phantom was 47.0 mm for the Siemens Mammomat Revelation. For the masses, the V i n t r a $V_{intra}$ for the intra-observer variation was 0.248, the averaged inter-observer variation, V ¯ i n t e r $\overline{V}_{inter}$ was 0.383. V p h a n t o m $V_{phantom}$ for phantom variance was 0.321. For the micro-calcifications, V i n t r a $V_{intra}$ was 0.0429, V ¯ i n t e r = $\overline{V}_{inter}=$ 0.0731 and V p h a n t o m = $V_{phantom}=$ 0.0759. CONCLUSIONS: Position, orientation and shape of the masses are reproducible and attenuation differences appropriate. The phantom presented proved to be a candidate test object for quality control.

Comparison of a personalized breast dosimetry method with standard dosimetry protocols.

Average glandular dose (AGD) in digital mammography crucially depends on the estimation of breast glandularity. In this study we compared three different methods of estimating glandularities according to Wu, Dance and Volpara with respect to resulting AGDs. Exposure data from 3050 patient images, acquired with a GE Senographe Essential constituted the study population of this work. We compared AGD (1) according to Dance et al. applying custom g, c, and s factors using HVL, breast thickness, patient age and incident air kerma (IAK) from the DICOM headers; (2) according to Wu et al. as determined by the GE system; and (3) AGD derived with the Dance model with personalized c factors using glandularity determined with the Volpara (Volpara Solutions, Wellington, New Zealand) software (Volpare AGD). The ratios of the resulting AGDs were analysed versus parameters influencing dose. The highest deviation between the resulting AGDs was found in the ratio of GE AGD to Volpara AGD for breast thicknesses between 20 and 40 mm (ratio: 0.80). For thicker breasts this ratio is close to one (1 ± 0.02 for breast thicknesses >60 mm). The Dance to Volpara ratio was between 0.86 (breast thickness 20-40 mm) and 0.99 (>80 mm), and Dance/GE AGD was between 1.07 (breast thickness 20-40 mm) and 0.98 (41-60, and >80 mm). Glandularities by Volpara were generally smaller than the one calculated with the Dance method. This effect is most pronounced for small breast thickness and older ages. Taking the considerable divergences between the AGDs from different methods into account, the selection of the method should by done carefully. As the Volpara method provides an analysis of the individual breast tissue, while the Wu and the Dance methods use look up tables and custom parameter sets, the Volpara method might be more appropriate if individual ADG values are sought. For regulatory purposes and comparison with diagnostic reference values, the method to be used needs to be defined exactly and clearly be stated. However, it should be accepted that dose values calculated with standardized models, like AGD and also effective dose, are afflicted with a considerable uncertainty budgets that need to be accounted for in the interpretation of these values.

Quality Controls in Digital Mammography protocol of the EFOMP Mammo Working group.

This article aims to present the protocol on Quality Controls in Digital Mammography published online in 2015 by the European Federation of Organisations for Medical Physics (EFOMP) which was developed by a Task Force under the Mammo Working Group. The main objective of this protocol was to define a minimum set of easily implemented quality control tests on digital mammography systems that can be used to assure the performance of a system within a set and acceptable range. Detailed step-by-step instructions have been provided, limiting as much as possible any misinterpretations or variations by the person performing. It is intended that these tests be implemented as part of the daily routine of medical physicists and system users throughout Europe in a harmonised way so allowing results to be compared. In this paper the main characteristics of the protocol are illustrated, including examples, together with a brief summary of the contents of each chapter. Finally, instructions for the download of the full protocol and of the related software tools are provided.

Conversion factors between human and automatic readouts of CDMAM phantom images of CR mammography systems.

In mammography screening, profound assessment of technical image quality is imperative. The European protocol for the quality control of the physical and technical aspects of mammography screening (EPQCM) suggests using an alternate fixed choice contrast-detail phantom-like CDMAM. For the evaluation of technical image quality, human or automated readouts can be used. For automatic evaluation, a software (cdcom) is provided by EUREF. If the automated readout indicates unacceptable image quality, additional human readout may be performed overriding the automated readout. As the latter systematically results in higher image quality ratings, conversion factors between both methods are regularly applied. Since most image quality issues with mammography systems arise within CR systems, an assessment restricted to CR systems with data from the Austrian Reference Center in the mammography screening program has been conducted. Forty-five CR systems were evaluated. Human readouts were performed with a randomisation software to avoid bias due to learning effects. Additional automatic evaluation allowed for the computation of conversion factors between human and automatic readouts. These factors were substantially lower compared to those suggested by EUREF, namely 1.21 compared to 1.62 (EUREF UK method) and 1.42 (EUREF EU method) for 0.1 mm, and 1.40 compared to 1.83 (EUREF UK) and 1.73 (EUREF EU) for 0.25 mm structure size, respectively. Using either of these factors to adjust patient dose in order to comply with image quality requirements results in differences in the dose increase of up to 90%. This necessitates a consensus on their proper application and limits the validity of the assessment methods. Clear criteria for CR systems based on appropriate studies should be promoted.

Spectrum optimization for computed radiography mammography systems.

PURPOSE: Technical quality assurance is a key issue in breast screening protocols. While full-field digital mammography systems produce excellent image quality at low dose, it appears difficult with computed radiography (CR) systems to fulfill the requirements for image quality, and to keep the dose below the limits. However, powder plate CR systems are still widely used, e.g., they represent ∼30% of the devices in the Austrian breast cancer screening program. For these systems the selection of an optimal spectrum is a key issue. METHODS: We investigated different anode/filter (A/F) combinations over the clinical range of tube voltages. The figure-of-merit (FOM) to be optimized was squared signal-difference-to-noise ratio divided by glandular dose. Measurements were performed on a Siemens Mammomat 3000 with a Fuji Profect reader (SiFu) and on a GE Senograph DMR with a Carestream reader (GECa). RESULTS: For 50mm PMMA the maximum FOM was found with a Mo/Rh spectrum between 27kVp and 29kVp, while with 60mm Mo/Rh at 28kVp (GECa) and W/Rh 25kVp (SiFu) were superior. For 70mm PMMA the Rh/Rh spectrum had a peak at about 31kVp (GECa). FOM increases from 10% to >100% are demonstrated. CONCLUSION: Optimization as proposed in this paper can either lead to dose reduction with comparable image quality or image quality improvement if necessary. For systems with limited A/F combinations the choice of tube voltage is of considerable importance. In this work, optimization of AEC parameters such as anode-filter combination and tube potential was demonstrated for mammographic CR systems.

A phantom using titanium and Landolt rings for image quality evaluation in mammography.

A phantom for image quality evaluation of digital mammography systems is presented and compared to the most widely used phantoms in Europe and the US. The phantom contains objects for subjective detection of Landolt rings (four-alternative, forced-choice task) and for objective calculation of signal-difference-to-noise ratios (SDNR), both in a titanium background within a 12-step wedge. Evaluating phantom images corresponding to exposures between 15 and 160 mAs (average glandular dose between 0.2 and 2 mGy), the resulting scores were compared to the scores obtained following the European EPQC and American College of Radiology (ACR) protocols. Scores of the Landolt test equal to 19 and 8.5 and SDNR equal to 20 and 11 were found to be equivalent to the acceptable limiting values suggested by EPQC and ACR. In addition, the Landolt and SDNR tests were shown to take into account the anatomical variations in thickness and tissue density within the breast. The simplified evaluation method presented was shown to be a sensitive, efficient and reliable alternative for image quality evaluation of mammography systems.

Teleradiology with uncompressed digital mammograms: clinical assessment.

PURPOSE: The purpose of our study was to demonstrate the feasibility of sending uncompressed digital mammograms in a teleradiologic setting without loss of information by comparing image quality, lesion detection, and BI-RADS assessment. MATERIALS AND METHODS: CDMAM phantoms were sent bidirectionally to two hospitals via the network. For the clinical aspect of the study, 200 patients were selected based on the BI-RAD system: 50% BI-RADS I and II; and 50% BI-RADS IV and V. Two hundred digital mammograms (800 views) were sent to two different institutions via a teleradiology network. Three readers evaluated those 200 mammography studies at institution 1 where the images originated, and in the two other institutions (institutions 2 and 3) where the images were sent. The readers assessed image quality, lesion detection, and BI-RADS classification. RESULTS: Automatic readout showed that CDMAM image quality was identical before and after transmission. The image quality of the 200 studies (total 600 mammograms) was rated as very good or good in 90-97% before and after transmission. Depending on the institution and the reader, only 2.5-9.5% of all studies were rated as poor. The congruence of the readers with respect to the final BI-RADS assessment ranged from 90% and 91% at institution 1 vs. institution 2, and from 86% to 92% at institution 1 vs. institution 3. The agreement was even higher for conformity of content (BI-RADS I or II and BI-RADS IV or V). Reader agreement in the three different institutions with regard to the detection of masses and calcifications, as well as BI-RADS classification, was very good (κ: 0.775-0.884). Results for interreader agreement were similar. CONCLUSION: Uncompressed digital mammograms can be transmitted to different institutions with different workstations, without loss of information. The transmission process does not significantly influence image quality, lesion detection, or BI-RADS rating.

Factors for conversion between human and automatic read-outs of CDMAM images.

PURPOSE: According to the European protocol for the quality control of the physical and technical aspects of mammography screening (EPQCM) image quality of digital mammography devices has to be assessed using human evaluation of the CDMAM contrast-detail phantom. This is accomplished by the determination of threshold thicknesses of gold disks with different diameters (0.08-2 mm) and revealed to be very time consuming. Therefore a software solution based on a nonprewhitening matched filter (NPW) model was developed at the University of Nijmegen. Factors for the conversion from automatic to human readouts have been determined by Young et al. [Proc. SPIE 614206, 1-13 (2006) and Proc. SPIE 6913, 69131C1 (2008)] using a huge amount of data of both human and automatic readouts. These factors depend on the observer groups and are purely phenomenological. The authors present an alternative approach to determine the factors by using the Rose observer model. METHODS: Their method uses the Rose theory which gives a relationship between threshold contrast, diameter of the object and number of incident photons. To estimate the conversion factors for the five diameters from 0.2 to 0.5 mm they exposed with five different current-time products which resulted in 25 equations with five unknowns. RESULTS: The theoretical conversion factors (in dependence of the diameters) amounted to be 1.61 ± 0.02 (0.2 mm diameter), 1.67 ± 0.02 (0.25 mm), 1.85 ± 0.02 (0.31 mm), 2.09 ± 0.02 (0.4 mm), and 2.28 ± 0.02 (0.5 mm). The corresponding phenomenological factors found in literature are 1.74 (0.2 mm), 1.78 (0.25 mm), 1.83 (0.31 mm), 1.88 (0.4 mm), and 1.93 (0.5 mm). CONCLUSIONS: They transferred the problem of determining the factors to a well known observer model which has been examined for many years and is also well established. This method reveals to be reproduceable and produces factors comparable to the phenomenological ones.

Validation of image quality in full-field digital mammography: is the replacement of wet by dry laser printers justified?

OBJECTIVE: Dry laser printers have replaced wet laser printers to produce hard copies of high-resolution digital images, primarily because of environmental concerns. However, no scientific research data have been published that compare the image quality of dry and wet laser printers in full-field digital mammography (FFDM). This study questions the image quality of these printers. MATERIALS AND METHODS: Objective image quality parameters of both printers were evaluated using a standardized printer test image, i.e., optical density and detectability of specific image elements (lines, curves, and shapes). Furthermore, mammograms of 129 patients with different breast tissue composition patterns were imaged with both printers. A total of 1806 subjective image quality parameters (brightness, contrast, and detail detection of anatomic structures), the detectability of breast lesions, as well as diagnostic performance according to the BI-RADS classification were evaluated. In addition, the presence of film artifacts was investigated. RESULTS: Optical density values were equal for the dry and the wet laser printer. Detection of specific image elements on the printer test image was not different. Ratings of subjective image quality parameters were equal, as were the detectability of breast lesions and the diagnostic performance. Dry laser printer images showed more artifacts (164 versus 27). However, these artifacts did not influence image quality. CONCLUSION: Based on the evidence of objective and subjective parameters, a dry laser printer equals the image quality of a wet laser printer in FFDM. Therefore, not only for reasons of environmental preference, the replacement of wet laser printers by dry laser printers in FFDM is justified.

Image quality of a wet laser printer versus a paper printer for full-field digital mammograms.

OBJECTIVE: The purpose of our study was to compare the image quality of a wet laser printer with that of a paper printer for full-field digital mammography (FFDM). MATERIALS AND METHODS: For both a wet laser printer and a paper printer connected to an FFDM system, image quality parameters were evaluated using a standardized printer test image (luminance density, dynamic range). The detectability of standardized objects on a phantom was also evaluated. Furthermore, 640 mammograms of 80 patients with different breast tissue composition patterns were imaged with both printers. Subjective image quality parameters (brightness, contrast, and detection of details of anatomic structures-that is, skin, subcutis, musculature, glandular tissue, and fat), the detectability of breast lesions (mass, calcifications), and the diagnostic performance according to the BI-RADS classification were evaluated. RESULTS: Both the luminance density and the dynamic range were superior for the wet laser printer. More standardized objects were visible on the phantom imaged with the wet laser printer than with the paper printer (13/16 vs 11/16). Each subjective image quality parameter of the mammograms from the wet laser printer was rated superior to those of the paper printer. Significantly more breast lesions were detected on the wet laser printer images than on the paper printer images (masses, 13 vs 10; calcifications, 65 vs 48; p < 0.05). With the paper printer images, BI-RADS 4 and 5 categories were underestimated for 10 (43.5%) of 23 patients. CONCLUSION: For FFDM, images obtained from a wet laser printer show superior objective and subjective image quality compared with a paper printer. As a consequence, the paper printer should not be used for FFDM.