Evaluation of breast amorphous calcifications by a computer-aided detection system in full-field digital mammography

Abstract

Objectives

The purpose of this study was to evaluate the performance of a direct computer-aided detection (d-CAD) system integrated with full-field digital mammography (FFDM) in assessment of amorphous calcifications.

Methods

From 1438 consecutive stereotactic-guided biopsies, FFDM images with amorphous calcifications were selected for retrospective evaluation by d-CAD in 122 females (mean age, 56 years; range, 35–84 years). The sensitivity, specificity, accuracy and false-positive rate of the d-CAD system were calculated in the total group of 124 lesions and in the subgroups based on breast density, mammographic lesion distribution and extension. Logistic regression analysis was used to stratify the risk of malignancy by patient risk factors and age.

Results

The d-CAD marked all (36/36) breast cancers, 85% (11/13) of the high-risk lesions and 80% (60/75) of benign amorphous calcifications (p<0.01) correctly. The sensitivity, specificity and diagnostic accuracy for the combined malignant and “high-risk” lesions was 96, 80 and 86%, respectively. The likelihood of malignancy was 29%. There was no significant difference between the marking of fatty or dense breasts (p>0.05); however, d-CAD marks showed differences for small (<7 mm) lesions (p=0.02) and clustered calcifications (p=0.03). The false-positive rate of d-CAD was 1.76 marks per full examination.

Conclusion

The d-CAD system correctly marked all biopsy-proven breast cancers and a large number of biopsy-proven high-risk lesions that presented as amorphous calcifications. Given our 29% likelihood of malignancy, imaging-guided biopsy appears to be a reasonable recommendation in cases of amorphous calcifications marked by d-CAD.

Many cancers that cannot be detected with mammography are in dense tissue, and digital mammography is more accurate in females with radiographically dense breasts [1]. In order for a radiologist to diagnose a breast cancer, this cancer must first be detected and second be correctly interpreted. Computer-aided detection (CAD) systems help radiologists with the perception of the cancer, marking regions of interest on the screen. The use of CAD software primarily in full-field digital mammography (FFDM), also known as direct CAD (d-CAD), does not require digitisation of the films and by definition, once an image is acquired, the CAD detection result will be reproducible when the same d-CAD scheme is applied repeatedly to such an image [2].

Breast microcalcifications are detected by high-quality mammography [3-8]. Amorphous calcifications are frequently difficult to identify on routine mammograms, with 78% seen retrospectively but not prospectively by Berg et al [9] in their series using screen-film mammograms. The purpose of CAD systems in breast imaging is to help the radiologist by marking suspicious regions that were initially missed; however, the radiologist has to decide whether true areas of concern are present in the highlighted locations and thus retains the ability to make the final decision about whether biopsy is deemed necessary.

Because amorphous calcifications without associated findings can be easily overlooked on routine screening mammograms, the use of d-CAD systems to aid in detection could be beneficial to the radiologist. Therefore, the purpose of our study was to determine the performance of one commercially available d-CAD system in marking non-palpable amorphous calcifications not associated with other mammographic findings and comparing this with the histopathology results.

Methods and materials

Institutional review board approval was obtained for this retrospective study. Informed patient consent was not required.

Study population

A database of 1438 consecutive stereotactic-guided biopsies that were performed between 2001 and 2007 was reviewed. Patients with FFDM (Senographe 2000D; GE Healthcare, Waukesha, WI) including four standard views and magnification views of the area of calcifications were identified. Cases with calcifications associated with a mass, distortion or palpable abnormality were excluded. Groups that contained calcifications meeting the description of branching, pleomorphic, punctate, round or dystrophic calcifications were also excluded from the study. Coincidental lesions other than amorphous calcifications seen on the mammograms were excluded from evaluation. Examinations meeting the inclusion criteria of the presence of prospectively reported amorphous calcifications without any other significant imaging finding were selected and reviewed by two independent radiologists (AMS, PC), both of whom were breast imaging fellowship-trained radiologists with more than 10 years of experience. In this study, the term amorphous includes microcalcifications that are “sufficiently small or hazy that a more specific morphologic classification cannot be made”, as defined by Breast Imaging Reporting and Data System [10]. There was an attempt to biopsy all the patients under stereotactic guidance and some of them required surgical excision for definitive diagnosis. The histopathology results served as truth for determining benign vs malignant results. This study comprised a total of 122 females (age range, 35–84 years; median, 56 years). Of these, 120 had unilateral findings and 2 had bilateral areas of subtle amorphous calcifications (Figure 1).

Figure 1.

Magnification view. A 68-year-old female with a cluster of amorphous microcalcifications in the left breast. Pathological findings at stereotactic biopsy revealed atypical ductal hyperplasia and columnar cell change associated with lobular involution and mild nuclear enlargement. Lumpectomy showed high-grade apocrine-type in situ duct carcinoma with focus of microinvasion.

Histopathology category

For statistical purposes, histopathology results were grouped into two categories: (1) benign lesions, for which surgical removal was not required; and (2) malignant-like lesions, which included malignancies, biopsies (graded according to the Nottingham classification scheme for cancer) [11] and “high-risk lesions” which required surgical excision for definitive diagnosis. Included in the high-risk lesion sub-group were findings of lobular neoplasia [atypical lobular hyperplasia (ALH) and lobular carcinoma in situ (LCIS)], atypical ductal hyperplasia (ADH), papillary lesions, flat epithelial atypia and radial scars.

Direct computer-aided detection system analysis

A commercially available d-CAD system (Second Look, v.7.2; iCAD, Nashua, NH) was applied retrospectively using a dedicated workstation (Advantage GE Workstation, GE Healthcare) in standard digital mammographic projections (craniocaudal and mediolateral oblique views) by a radiologist (AMS) fellowship trained in breast imaging, who had access to the surgical database and was able to correlate the precise location of d-CAD marks with the location of pre-operative needle (wire) localisation in patients who underwent surgical excision. This person recorded all mammographic descriptors: American College of Radiology (ACR) mammographic density, extent (maximum extension of the area of amorphous calcifications), distribution (clustered, multiple clusters, regional, linear, segmental) and patient risk factors (age, family history, prior or concurrent atypical hyperplasia or cancer). The same study radiologist also assessed the rates of malignancy for each subgroup.

True-positive and false-negative d-CAD marks

If the d-CAD system marked on at least one view the mammographic location of the cancer or high-risk lesion, the case was considered a true-positive d-CAD mark. Malignant lesions or high-risk lesions that were not marked were considered a false-negative d-CAD mark.

False-positive d-CAD marks and distractive d-CAD marks (false-positive rate)

The distractive (false-positive rate) marks consist of images with ellipses and rectangles highlighting potential areas of concern not related to the area biopsied. The ellipses mark potential masses (non-spiculated masses, spiculated masses, architectural distortions or asymmetric densities) and the rectangles mark potential microcalcifications. The false-positive rate was calculated by considering all distractive d-CAD marks (masses and calcifications) displayed on the screen not related to the biopsied area of amorphous calcification. False-positive d-CAD marks were considered to be only those marks that correctly indicated the location of the amorphous calcifications that proved not to be cancers or high-risk lesions.

Full-field digital mammography follow-up

Mammographic follow-up was performed annually for all patients who did not undergo a mastectomy. d-CAD was applied to all available cases to be certain that the same marked area would not develop a cancer within a minimum of 1 year [FFDM mean (standard deviation; SD) follow-up time was 3.73±1.33 years]. This was used to assess d-CAD sensitivity, specificity, negative predictive value, positive predictive value and accuracy.

Statistical analysis

Contingency tables were analysed using the χ² or Fisher exact test to determine whether any significant difference existed between the benign and malignant categories based on size, density and distribution. Logistic regression analysis was performed to estimate the effect of patient risk factors—age, prior breast cancer or prior high-risk lesions, and family history of breast cancer, on the probability of malignancy or a high-risk lesion.

For statistical purposes breast density was grouped into only two categories: fatty breast tissue (ACR density 1 and 2) or dense breast tissue (ACR density 3 and 4).

Results

The d-CAD system correctly marked 100% (36/36) of the malignant microcalcifications and 85% (11/13) of the high-risk lesions. The final histopathology results are summarised in Table 1. In 17 cases (15/75 benign and 2/13 high-risk lesions), the amorphous microcalcifications (14%) were not marked by d-CAD and none proved to be malignant after a mean (SD) time of 3.67±0.89 years of FFDM follow-up.

Table 1. Summary of the final histopathological findings for 124 lesions that manifested as amorphous calcifications.

Finding	Number of cases	Percentage
Malignant lesions	36	29
Invasive and in situ carcinoma	5	14
High-grade DCIS with microinvasion	3	8
High-grade DCIS	9	25
High-grade DCIS with LCIS	1	3
High- and intermediate-grade DCIS	4	11
Intermediate-grade DCIS	10	28
Low/intermediate-grade DCIS	2	6
Low-grade DCIS with LCIS	2	6
High-risk lesions	13	10
Atypical duct hyperplasia	6	46
Atypical lobular hyperplasia	3	23
Flat epithelial atypia	3	23
LCIS	1	8
Benign lesions	75	60
Fibrocystic changes	27	36
Benign stromal calcifications	13	17
Usual duct hyperplasia	9	12
Columnar cell change without atypia	6	8
Fibrosis	5	7
Fibroadenomatoid foci with microcalcifications	4	5
Sclerosing adenosis	3	4
Intraductal papilloma	2	3
Mild chronic mastitis	1	1
Biopsy not performeda	5	7

DCIS, ductal carcinoma in situ; LCIS, lobular carcinoma in situ.

^aFollow-up mean (standard deviation) 3.35 (2.63–4.84) years.

We found a significant statistical difference between the malignant-like group (47/49) and benign group (60/75) marked by d-CAD (p<0.05). The positive predictive value was 76% and the negative predictive value was 97%. Sensitivity was 96%, specificity was 80% and diagnostic accuracy was 86% (Table 2).

Table 2. Case-based performance of direct computer-aided detection (d-CAD) system of amorphous calcifications based on mammographic density.

Calcifications	d-CAD marks	Sensitivity/specificity/accuracy (%)	p-value
Densea malignant + high risk	27/29
Densea benign	34/40	93/85/88	0.001
Fattyb malignant + high risk	20/20
Fattyb benign	26/35	100/74/84	0.001
All malignant + high risk	47/49
All benign	60/75	96/39/86	0.0147

Table data are the number of amorphous calcification lesions marked by d-CAD out of the number of calcification lesions in the category.

^aDense calcifications are full-field digital mammography (FFDM) with ACR mammographic density 3 and 4.

^bFatty calcifications are FFDM with ACR mammographic density 1 and 2.

Breast density

Table 2 shows the performance of the d-CAD system related to breast density. The sensitivity of the d-CAD system for malignant-like lesions in fatty breasts was 100% (20/20) and was 93% (27/29) in dense breasts. Specificity was 74% for fatty breast tissue and 85% for dense breast tissue. There was no significant difference (p>0.1) in d-CAD marks between dense and fatty breasts to differentiate benign from malignant-like groups.

Lesion dimensions and distribution

The extension of the area of amorphous calcifications was 7.5 mm (range, 1–44 mm). Calcification distribution was found to be clustered (n=108), segmental (n=6), regional (n=5) or linear (n=5). Malignant-like groups of amorphous calcifications ranged from 2 to 29 mm in the greatest extent (mean, 8 mm) and were clustered (n=39), linear (n=3), segmental (n=3) or regional (n=4). Benign lesions ranged from 1 to 44 mm (mean, 6 mm) and were clustered (n=69), segmental (n=3), linear (n=2) or regional (n=1) in distribution. Clustered amorphous calcifications showed a significant difference of d-CAD marks (p=0.03) between the benign and malignant-like groups (Table 3). Other distributions did not show a significant difference. The d-CAD tended to mark smaller (in extent) malignant calcifications than those with a similar extension in the benign group (p=0.02).

Table 3. Case-based performance of direct computer-aided detection (d-CAD) system of amorphous calcifications based on mammographic extension and distribution.

Calcifications	d-CAD marks	Sensitivity/specificity/accuracy (%)	p-value
1–7 mm malignant/high risk	27/28
1–7 mm benign	41/55	96/75/82	0.001
>7 mm malignant/high risk	20/21
>7 mm benign	19/20	95/95/95	>0.001
Clustered malignant	37/39
Clustered benign	54/69	94/78/84	0.0002

Table data are the number of amorphous calcification lesions marked by d-CAD out of the number of calcification lesions in the category.

Patient risk factors

A trend towards an increased risk for malignancy [relative risk (RR)=1.76 (range, 0.76–4.04)] was seen in patients who had a prior breast biopsy showing ADH or LCIS, or a personal history of breast cancer; however, the trend was not statistically significant. The logistic regression analysis of family history of breast cancer also indicated that individuals with a first-degree relative with breast cancer had 1.2683 greater odds of developing breast cancer than individuals with amorphous calcifications without a family history of breast cancer. Table 4 shows the likelihood of outcomes of malignant lesions, high-risk lesions and benign calcifications as a function of the patient's age. The oldest age group, 75–84 years, showed a higher risk of malignant outcome with an odds ratio of 1.49 (95% confidence interval: 0.66, 3.37), although there was a small number of such lesions and patients.

Table 4. Likelihood of histopathological findings as a function of patient's age.

Age (years)	Number of lesions	Number of malignancies (%)	Number of high-risk lesions (%)	Number of benign lesion (%)	p-value
35–44	13	1 (8)	0 (0)	12 (92)	0.99
45–54	41	15 (37)	2 (5)	24 (59)	0.72
55–64	42	12 (29)	8 (19)	22 (52)	0.81
65–74	22	5 (23)	2 (9)	15 (68)	0.54
75–84	6	3 (50)	1 (17)	2 (33)	0.33
Total	124	36 (29)	13 (10)	75 (60)	NA

NA, not applicable.

Table data are the number of amorphous calcification lesions in each category.

False-positive marks

Mammographic long-term follow-up with annual FFDM was available for 109 of the 122 patients included in the study. Of the 36 malignant cases, 6 patients had mastectomy, 26 patients had lumpectomy and 4 patients had surgical follow-up outside our institution. Patients treated with a lumpectomy did not have additional cancers with in a mean time of 4.04 years (range, 1.01–7.24 years) of mammographic follow-up. None of the 13 patients with the high-risk lesions missed the imaging follow-up and they were free of new d-CAD marks for calcifications at 3.3 years of follow-up. Of the 75 patients with benign biopsy or lumpectomy results, just 1 patient was lost to FFDM follow-up. A chart review revealed that this particular patient, who had a history of lung carcinoma and brain metastasis, had a breast MRI carried out 2 years after the benign breast biopsy with negative results. None of the 74 patients with FFDM follow-up developed interval cancers after 3.79 years mean imaging follow-up time. The false-positive d-CAD was 80% (60/75) lesions.

In addition, the d-CAD system displayed 222 distracting false marks (98 calcification marks and 124 mass marks) in both breasts, with a mean of 1.76 false-positive marks per 4-image case. These CAD marks were not considered as false-positives in the calculation of d-CAD sensitivity, specificity or accuracy for depiction of malignant amorphous calcifications, and this value was used only as the false-positive rate of the software.

Discussion

The results of our study demonstrate that the d-CAD system is highly sensitive in detecting amorphous calcifications which proved to be malignant or high-risk lesions. We found a 29% risk of malignancy of the amorphous calcifications, and d-CAD detected 100% of the malignant cases or 96% of the malignant-like cases (breast cancer plus high-risk lesions). Berg et al [9] and Soo et al [12] found a 16–20% risk of malignancy in their series, and CAD case sensitivities by histological outcomes were found to be 57% for malignant calcifications, 29% for high-risk calcifications and 54% for benign calcifications. Their results are not akin to ours (100% for malignant calcifications, 85% for high-risk calcifications and 61% for benign calcifications) and this may represent the differences from using direct (our study) vs indirect CAD (with digitisation of images) systems. In the field of CAD, the issue of reliability and reproducibility, which has been shown to be far from perfect in screen-film mammography (SFM) [2,13,14] as a result of mechanical scanning, virtually does not exist in FFDM since the mechanical process of scanning is avoided and CAD software is directly applied to the digital mammogram.

Amorphous calcifications are frequently overlooked on SFM and can contribute to a substantial number of missed carcinomas [9,12,15,16]. Logically, the use of CAD systems that have superb sensitivity for the detection of malignant microcalcifications seems to be very attractive in digital mammography. Previous studies [12,17-19] have shown that CAD sensitivity for the detection of breast cancer is similar in fatty breasts and dense breasts in both SFM [17] and FFDM [20-23]; however, indirect-CAD sensitivity for subgroups of cancers apparent only as microcalcifications has been shown to be higher in dense breasts than the sensitivity for calcifications in fatty breasts [12]. The indirect-CAD sensitivity was markedly lower in cases of malignant amorphous calcifications than in other malignant calcifications [12]. In our study, a similar trend was seen whereby d-CAD did not miss any cancers among fatty breasts; however, there were no statistically significant differences between dense and fatty breasts to correctly mark the malignant-like lesions. This may be related to the fact that, despite improved CAD algorithms for the detection of malignancy in dense breast tissue and the inherent advantages of FFDM in detection of cancers in dense breast tissue [1], amorphous calcifications are frequently at the threshold of visibility for radiologists and detectability for current algorithms used by vendors.

Early detection of small cancers on screening examinations is important. In our study, d-CAD marked 96% of the malignant-like group presenting as amorphous calcifications measuring 1–7 mm extent and 95% of the malignant-like group measuring >7 mm (Table 3), which is higher than published results for d-CAD [22] for this size of group, and marked higher than indirect CAD applied in amorphous calcifications [12] or slightly higher when considering all types of microcalcifications [24] for equivalent extension measurements.

Patient risk factors affect the likelihood of malignancy [25]. Berg et al [9] observed that biopsies of amorphous calcifications in the same breast as a past or current cancer were more likely to be malignant (p=0.025) and found a trend towards and increased likelihood of malignancy with a prior or concurrent history of a high-risk lesion or cancer. Although our results failed to reach statistical significance, we observed that, in females with cancer or prior breast biopsy showing ADH, contralateral breast biopsies were nearly twice (RR=1.76) as likely to be malignant or show high-risk lesions. We did not include ADH as a benign result and decided to include it in the group named malignant-like because recent studies that compared benign tissue, ADH, DCIS and invasive cancer, using mRNA expression profiling, quantitative real-time polymerase chain reaction analysis and analysis of allelic imbalance, showed that ADH is a genetically advanced pre-cancerous lesion and that ADH, similar to DCIS, is a non-obligatory direct precursor of invasive ductal cancer [26-29]. This emphasises the need for early detection of pre-cancerous lesions, particularly among amorphous calcifications. Breast surgeons consider the removal of a lesion containing ADH as standard treatment/secondary prevention because surgical resection without chemotherapy or radiotherapy has been extremely effective in reducing the risk of recurrence [28,29]. In addition, our results proved that amorphous calcifications can be low-to-intermediate grade DCIS. Therefore, identifying and biopsying amorphous calcifications is important, and d-CAD potentially can be very helpful in detecting these calcifications.

Knowledge of false-positive marks generated by d-CAD, not related to the region of interest, is important because screen pollution with false marks can potentially distract the radiologist's interpretation. In our study, 1.76 false-positive marks were observed in examinations with four standard views. This is lower than the 2.8–5.2 marks per case achieved using indirect CAD [15,17,30,31] and lower than a recent d-CAD study [22] using the same software version, which showed 2.3 marks per case. In another recent retrospective study of CAD with FFDM, Yang et al [21] found a similar false-positive rate of 1.8 per patient case using a different d-CAD system.

The limitations of our study include its retrospective design and lack of assessment of d-CAD effect on workflow and radiologists' performance. In addition, we did not have imaging follow-up for 1 of the 75 patients with benign biopsy/benign lumpectomy results included in the false-positive analysis. For this reason, we cannot confirm whether the marks considered as false positive may have been true-positive d-CAD marks.

Successful d-CAD detection of all malignant calcifications in our study may need further validation with prospective studies before the possible future introduction in the algorithm of management of amorphous calcifications. However, since amorphous calcifications would be particularly difficult lesions to follow, given the difficulty in counting each small, indistinct, very faint calcified particle (Figure 1), and since accurately determining whether the number of calcifications had increased during the follow-up interval would be extremely difficult, we suspect that waiting for more highly suspicious changes, such as development of pleomorphic or linear branching calcifications or even development of an associated mass, raises the concern of a worsened prognosis for the patient; for this reason, imaging-guided biopsy appears reasonable to be recommended in cases of amorphous calcifications marked by d-CAD given our 29% likelihood of malignancy, which includes pure high-grade DCIS, DCIS with microinvasion and invasive carcinoma. Validation of this approach across multiple centres is needed.