Can occult invasive disease in ductal carcinoma in situ be predicted using computer-extracted mammographic features?

Abstract

Rationale and Objectives

This study aimed to determine whether mammographic features assessed by radiologists and using computer algorithms are prognostic of occult invasive disease for patients showing ductal carcinoma in situ (DCIS) only in core biopsy.

Materials and Methods

In this retrospective study, we analyzed data from 99 subjects with DCIS (74 pure DCIS; 25 DCIS with occult invasion). We developed a computer-vision algorithm capable of extracting 113 features from magnification views in mammograms and combining these features to predict whether a DCIS case will be upstaged to invasive cancer at the time of definitive surgery. In comparison, we also built predictive models based on physician-interpreted features, which included histologic features extracted from biopsy reports and Breast Imaging Reporting and Data Systems (BI-RADS) related mammographic features assessed by two radiologists. The generalization performance was assessed using leave one out cross validation with the receiver operating characteristic (ROC) curve analysis.

Results

Using the computer-extracted mammographic features, the multivariate classifier was able to distinguish DCIS with occult invasion from pure DCIS, with an area under the curve for ROC (AUC-ROC) equal to 0.70 (95% CI: 0.59–0.81). The physician-interpreted features including histologic features and BI-RADS related mammographic features assessed by two radiologists showed mixed results, and only one radiologist’s subjective assessment was predictive, with AUC-ROC equal to 0.68 (95% CI: 0.57–0.81).

Conclusion

Predicting upstaging for DCIS based upon mammograms is challenging, and there exists significant inter-observer variability among radiologists. However, the proposed computer-extracted mammographic features are promising for the prediction of occult invasion in DCIS.

Introduction

Ductal carcinoma in situ (DCIS) is a preinvasive tumor confined within the ducts of the mammary glands [1], and it lies along the breast cancer continuum between atypical ductal hyperplasia and invasive ductal carcinoma. The incidence of DCIS has increased substantially since the introduction of mammographic screening with over 60,000 women in the United States diagnosed with DCIS every year, representing approximately 20% of all new breast neoplasm diagnoses [2]. However, despite the increased incidence of DCIS there has not been a concomitant decrease in invasive breast cancer [3]. Since the risk of progression from DCIS to invasive cancer is unclear, with estimates ranging from 14% to 53% [4], there is a growing debate about overdiagnosis and consequent overtreatment of DCIS. Furthermore, among DCIS-only cases diagnosed at core biopsy, approximately 26% will be shown to contain invasive ductal carcinoma at surgical excision [5]. This upstaging, specifically from DCIS diagnosed at core biopsy to invasive ductal carcinoma at excision, has important consequences for patient management.

Many studies have sought to predict the occult invasion in DCIS. Different factors or markers, including immunohistochemical biomarkers, histological features, and mammographic or sonographic findings, have been described and associated with outcomes in DCIS [5–11]. However, none of these factors have been accepted as a definitive predictor of this upstaging or is sufficiently reliable for clinical use. Overall, it still remains a difficult task and unmet need to accurately predict occult invasive disease in DCIS.

Breast microcalcifications (MCs) appear in 30–50% of mammographically detected cancers [12], and over 90% of women with DCIS have suspicious MCs on mammography [13]. There has been much work using computer-aided detection (CAD) and diagnosis (CADx) for mammography including microcalcification clusters [14–20]. Other CAD/CADx studies have focused on DCIS [21–24], but those studies have not utilized the diagnostic magnification views routinely available during the workup of suspicious calcifications, which offer additional details not appreciable on routine full-field screening mammographic views. In this work, we hypothesize that computer vision techniques as well as various mammographic features developed for screening detection or diagnosis can be used to help predict the presence of occult invasive disease associated with DCIS. We have, therefore, developed a computer-vision algorithm based approach to extract mammographic features for patients with DCIS, and built a classification model relying on these features to distinguish between pure DCIS and DCIS with occult invasive disease.

Materials and Methods

The processing pipeline to analyze the digital mammography magnification views is shown in Fig. 1. A core component of our prediction model is a computer-vision algorithm based approach to extract mammographic features (denoted as computer vision features in the remaining sections). In addition, we also extracted and evaluated some histologic features and BI-RADS (Breast Imaging Reporting and Data Systems) related mammographic features that have been used in the literature [1, 5, 6, 25–29](denoted as physician-interpreted features in the remaining sections), and compared their predictive power for occult invasive disease in DCIS with the computer vision features.

Figure 1.

Flowchart of the proposed methodology.

Subject selection

This study was approved by the institutional review board of XXX. All procedures were compliant with the Health Insurance Portability and Accountability Act. Our subjects included women age 40 and older who underwent stereotactic core needle biopsy and were diagnosed with pure DCIS that presented as calcifications only, and for whom at least one digital magnification view was available. Exclusion criteria included the presence of any masses, asymmetries, or architectural distortion on a mammogram; history of breast cancer or prior surgery; and presence of microinvasion at the time of initial biopsy. The excluded subjects were deemed to be already at high risk for invasion and therefore not appropriate for this study. Overall, 99 subjects from our institution with biopsy-proven DCIS only were retrieved from 2009–2014. Of those, 25 were upstaged to invasive cancer at the time of definitive surgery. All magnification views were produced by GE Senographe Essential systems, with either 1.5X or 1.8X magnification.

Detection of individual microcalcifications and clusters

A computer-vision algorithm based approach was used to segment individual MCs and clusters within the DCIS lesion followed by extraction of mammographic features from the segmented areas to predict the upstaging of DCIS. To accomplish this, calcification regions of interest (ROIs) were first identified by an expert radiologist. Following this manual step, we applied an automatic algorithm described below:

1. Mammogram enhancement:

   We first enhanced the detectability of MCs through contrast-limited adaptive histogram equalization with a 1mm by 1mm local tile size. Then, a dual-structural element based morphological operator followed by a top-hat transform [20] was applied to further suppress non-relevant background noise. Fig. 2 shows the process of this enhancement procedure.

2. Coarse detection of microcalcifications:

   Calcifications occur as bright points in a complex background with heterogeneous intensity and texture. Thus, we proposed a locally adaptive, triple-thresholding technique to detect the potential microcalcification candidates. Considering the size of MCs, we defined a 1mm by 1mm local window. Then, a local intensity mean image F_{local -mean} and a local intensity standard deviation image F_local-STD were constructed from the previously enhanced image F by assigning each pixel with the average intensity value and intensity standard deviation computed within the local window. Next, we conducted histogram thresholding independently on these three images F, F_{local -mean}, F_local-STD, with top 1%, 2% and 2% pixels kept respectively. All three thresholdings were applied within the DCIS ROI area pre-delineated by an expert breast radiologist. In the end, the overlaps of the resulted three thresholded areas were both globally and locally bright small spots, which were regarded as detected MC candidates.

3. False positive reduction:

   We further eliminated the false positives from the MC candidates based on their morphologic characteristics. Candidates were removed if they were smaller than 0.1 mm² or larger than 4 mm²; had circularity within the range 0.9 – 1.1 and area larger than 1mm², (large, obviously benign calcifications); or had axis ratio larger than 5 (rod-like shapes). By removing those false positives, we obtained the final refined segmentation for individual MCs.

4. Detection of cluster boundary:

   Starting from the center of each detected MCs, a weighted graph was built by connecting all the pairs of MCs that are less than 10 mm from each other and removing the isolated ones. The boundary of the cluster was generated using the convex hull algorithm. Fig. 3 shows the final segmentation results for the same example as in Fig. 2.

Figure 2.

Result of mammogram enhancement: (a) the original digital magnification view; (b) enhanced image by contrast-limited adaptive histogram equalization and dual-structural element based morphology approach; (c) final enhanced image after applying top-hat transform.

Figure 3.

Example result of segmentation of individual MCs and detection of cluster boundary: (a) DCIS ROI mask delineated by radiologist; (b) segmented by the algorithm.

Feature extraction

Computer vision features

For this task, we built a comprehensive feature set similar to that described by Bria et al. [14]. Specifically, from the segmented MCs and clusters, we extracted three types of mammographic features: (i) shape features to describe heterogeneous morphology and size of MC and clusters; (ii) topological features capturing the topological relations between MCs from the weighted graphs associated with the clusters; (iii) texture features obtainable from the original magnification views, such as MC pixel values, background values and statistical measures of gray level co-occurrence matrices (GLCM). Individual MC-level features were computed for all MCs belonging to the same cluster. From these values, four global statistical measures (Mean, STD, Min, and Max) were computed to describe the cluster. For each cluster, we obtained 113 features: 100 individual MC-level and 13 cluster-level features, which are shown in Table. 1. We denote this feature set as computer vision features. For more detailed description of these features, we refer readers to [14].

Table 1.

Computer vision features extracted from individual MCs and cluster.

	Category	Feature	Description
Individual MCs	Shape	MC perimeter (p)	Length of MC contour
		MC area (Amc)	Number of pixels for a MC × pixel area (µm2)
		MC circularity	4π*Amc/p2, a measure of “roundness”
		MC eccentricity	Another measure of “roundness”
		MC major axis	Length of major axis of MC region
		MC minor axis	Length of minor axis of MC region
		MC Hu’s moments *7	Descriptive weighted averages of intensities
	Topological	MC distance2centroid	Distance to cluster centroid
		MC distance2closest	Distance to nearest MC neighbor
		MC degree	Number of edges incident to MC
		MC normalized degree	Sum of normalized weights of MC degree
	Texture	MC background *2	Mean and std. of background pixel intensities
		MC foreground *2	Mean and std. of MCs’ pixel intensities
		MC GLCM * 4	Measures computed from GLCMs
Cluster	Shape	MCCs area (Ac)	Area of cluster
		MCCs eccentricity	Eccentricity of cluster
	Topological	MCCs number (n)	Number of MCs in cluster
		MCCs density	2E/n*(n-1): E is number of graph edges
		MCCs coverage	sum(Amc/Ac)
	Texture	Cluster background *2	Mean and std. of background pixel intensities in cluster
		Cluster foreground *2	Mean and std. of MCs’ pixel intensities in cluster
		Cluster GLCM * 4	Measures computed from GLCMs for whole cluster

Physician-interpreted features

Prior studies suggested relationships between outcomes of DCIS and histologic features from pathology reports [1, 5, 6, 27, 29] or BI-RADS mammographic features [5, 25–28]. Although not intended for the present task, those features nonetheless may have predictive value. Therefore, we recorded another set of features (denoted as physician-interpreted features) to compare to the computer vision features. From the core biopsy pathology reports, two histologic features were identified: the nuclear grade (low, intermediate, or high) and the presence or absence of comedo type necrosis. Each case was reviewed independently by two fellowship-trained, dedicated breast radiologists, who were blinded to whether invasive disease was eventually found. One of the radiologists (denoted as Radiologist-A in the remaining sections) has 3 years of experience in breast imaging; the other radiologist (denoted as Radiologist-B) has 20 years of experience in breast imaging. By following the terminology of BI-RADS, the following mammographic features were obtained independently from two radiologists: calcification description, distribution, and BI-RADS assessment (4a, 4b, 4c, and 5). Since radiologists were aware that all included subjects had biopsy-proven DCIS, their BI-RADS assessments should all be within category 4a, 4b, 4c, or 5. In addition, we recorded the patient’s age and lesion size as well. The lesion size was measured from each radiologist’s outline of the lesion on the best magnification view by both 2D area and the major axis length. For the description and distribution of calcifications, the options were further ranked with the order of increasing risk of being abnormal, similar to [30]. Furthermore, the radiologists were also asked to provide another ranking score (between 1–100) as their overall judgment of the probability of occult invasion. In total, nine different types of physician-interpreted features were extracted, as shown in Table 2.

Table 2.

Comparison of histologic and mammographic features of the DCIS and invasive groups.

Feature		DCIS (74)	Invasion (25)	p-value
Age		59.8	58.2	p ≤ 0.5330
Histologic	Nuclear grade (1–3)	2.51	2.58	p ≤ 0.6044
	1	6	3
	2	27	6
	3	41	16
	Subtype of DCIS			p ≤ 0.5250
	Comedo	36	14
	Non-Comedo	38	11
		Radiologist-A			Radiologist-B
Mammographic	Size of lesion
	Area (mm^2)	210.3	369.0	p ≤ 0.2257	182.5	442.9	p ≤ 0.0305 *
	Major Axis (mm)	16.7	24.8	p ≤ 0.0496 *	17.3	25.4	p ≤ 0.0345 *
	Morphology of calcifications			p ≤ 0.0704			p ≤ 0.8690
	Low Risk (Typically benign)	0	1		0	0
	Medium Risk (Amorphous/Coarse heterogeneous)	41	7		16	5
	High Risk (Fine pleomorphic/Fine linear)	33	17		58	20
	Distribution of calcifications			p ≤ 0.6653			p ≤ 0.2056
	Regional	2	0		11	3
	Segmental	5	3		3	4
	Linear	2	1		15	7
	Clustered	65	21		45	11
	BI-RADS level of suspicion			p ≤ 0.0247 *			p ≤ 0.0321*
	4a	40	7		1	1
	4b	17	8		35	6
	4c	14	8		23	7
	5	3	2		15	11
	Radiologist’s score of being invasive	14.5	21.0	p ≤ 0.0052 *	32.2	35.4	p ≤ 0.4965

^*Difference for the comparison was statistically significant.

Building and evaluating predictive models

We developed predictive models for both computer vision features as well as physician-interpreted features using univariate and multivariate analyses.

To assess the performance of individual computer vision features, we directly used their normalized numerical values as the posterior probability of being invasive. In the multivariate setting, a logistic regression classifier was utilized with sequential floating forward feature selection (SFFS) [31] nested within leave-one-out cross validation (LOOCV).

For the physician-interpreted features, in univariate setting we performed the Student’s t-test for the continuous features (i.e., age, size of lesion), chi-squared analysis for the nominal feature (i.e., subtype of DCIS), and Mann-Whitney U-test for the other ordinal features. The significance level was 0.05. For multivariate analysis, all features were converted into numeric form (0–1) and then modeled with logistic regression with LOOCV.

The generalization performance of the predictive model was assessed using ROC analysis. To analyze whether the area under the curve of different ROC curves differ significantly, comparisons of AUCs were performed using DeLong’s method [32] implemented in the pROC package [33] with R (V.3.3.1).

Results

Univariate performance

Computer vision features

The AUC-ROC performance for individual computer vision features is shown in Fig. 4. Overall, the AUC-ROC performance averaged over all 113 computer vision features is 0.58 ± 0.06, with 25 out of 113 features performing better than chance (p<0.05) at predicting occult invasive disease. Specifically, for individual cluster-wise features (features indexed between 1–13 in Fig.4), the averaged AUC-ROC performance cluster-wise features is 0.61 ± 0.07, with 7 out of 13 features in this group performing better than chance (p<0.05); for individual MC-level features using “mean” statistic (features indexed between 14–38 in Fig.4), the averaged AUC-ROC performance is 0.54 ± 0.05, with 3 out of 25 features in this group performing better than chance (p<0.05); 0.58 ± 0.05 for individual MC-level features using “standard deviation” statistic (features indexed between 39–63 in Fig.4), with 5 features performing better than chance (p<0.05); 0.57 ± 0.04 for individual MC-level features using “minimum” statistic (features indexed between 64–88 in Fig.4), with 3 features performing better than chance (p<0.05); 0.60 ± 0.04 for “maximum”, with 7 features performing better than chance (p<0.05).

Figure 4.

The AUC-ROC performance of individual computer vision features. The red line indicates chance behavior for AUC-ROC being 0.5. The feature groups are cluster level (black) and summary statistics of individual calcification features: mean (green), standard deviation (blue), minimum (yellow), and maximum (cyan).

Physician-interpreted features

Results of the statistical analysis on physician-interpreted features are presented in Table 2. Between the pure DCIS and the invasive groups, there was no statistically significant difference in patient age and histologic features, i.e., nuclear grade, subtype of DCIS. For both radiologists, the morphology and distribution of calcifications showed no statistically significant difference between the pure DCIS and the invasive groups. However, the major axis length of DCIS lesion and BI-RADS levels of suspicion reached statistical significance for both radiologists (Radiologist-A: p ≤ 0.0496, p ≤ 0.0247; Radiologist-B: p ≤ 0.0345, p ≤ 0.0321). The lesion area was significant for only Radiologist-B (p ≤ 0.0305). The radiologist’s subjective score of being invasive was significant for only Radiologist-A (p ≤ 0.0052). Since this unique feature is directly associated with our specific task, we further assessed the inter-observer variability using intraclass correlation coefficients (ICC) (1-way model, consistency definition). The ICC showed a fair correlation of 0.53 (95% CI: 0.37–0.66) between the two radiologists.

Multivariate performance

Fig. 5 shows the ROC curves of the proposed predictive model using different features. For the physician-interpreted features, given that just Radiologist-A’s assessment was significant, only his BI-RADS related features were utilized. Using that subjective assessment as the posterior probability of being invasive achieved a univariate AUC-ROC of 0.68 (95% CI: 0.57–0.81), which performed significantly better than random chance with AUC-ROC of 0.5 (p<0.05) at predicting occult invasive disease. However, in multivariate modeling, building a logistic regression classifier by combining Radiologist-A’s three statistically significant features as shown in Table 2 only achieved AUC-ROC of 0.59 (95% CI: 0.45–0.73), compared to 0.52 (95% CI: 0.38–0.67) by combining all nine physician-interpreted features. Both of these two multivariate predictive models failed to reject the null hypothesis of random chance with AUC-ROC of 0.5.

Figure 5.

ROC curves showing the classification performance using different features. The ROC curve for Radiologist-B’s subjective assessment achieved AUC-ROC of 0.55 (95% CI: 0.41–0.68) and showed no statistically significant difference from random chance and was thus not plotted.

By using the computer vision features, the proposed multivariate predictive model of using sequential floating forward feature selection (SFFS) and logistic regression classifier achieved AUC-ROC of 0.70 (95% CI: 0.59–0.81). Note that this corresponds to the best of the models that we implemented in our analysis; more details are included in the discussion. This multivariate model based on computer vision features significantly outperformed random chance (p<0.05), but shows no statistically significant difference with the univariate model of Radiologist-A’s subjective assessment (p>0.05).

We also plotted the histogram of feature selection frequency when using SFFS method, as shown in Fig. 6. The selection frequencies for the 113 computer vision features were obtained by normalizing the overall 99 folds in LOOCV. The features among the top picks are associated with certain textural properties within clusters (features indexed at 63, 59, 37), the MC density of the cluster (feature indexed at 29), and the size and shape of the MCs (features indexed at 42 and 65).

Figure 6.

The histogram of feature selection frequency using computer vision features and SFFS across the cross validation. These feature groups are cluster level (black) and summary statistics of individual calcification features: mean (green), standard deviation (blue), minimum (yellow), and maximum (cyan).

Discussion

Our study is the first of its kind to examine the feasibility of using computer-extracted mammographic features to predict occult invasive disease in patients with core needle biopsy proven DCIS that presents as mammographic calcifications. In univariate analysis, 25 computer vision features out of total 113 showed significantly better-than-chance predictive power for detecting occult invasive disease. In comparison, there were mixed results among the radiologist interpreted features: the subjective assessments by two radiologists led to different results, with only one radiologist’s subjective score being predictive, although the DCIS lesion size measured in the major axis and the BI-RADS level of suspicion assessed by both radiologists showed statistically significant difference between the pure DCIS and the invasive groups. On multivariate modeling, the computer vision features significantly outperformed random chance (p<0.05), but with an AUC-ROC of 0.70 that showed only marginal improvements over Radiologist-A’s subjective assessment. This emphasizes the difficult nature of predicting the presence of invasive cancer in this setting.

In this initial exploratory analysis, we applied computer-vision features previously utilized in CAD systems, which were designed primarily for the detection of cancer [34–38]. While a useful starting point, the task of identifying a suspicious abnormality is different from discriminating between known DCIS and occult invasive disease. This likely explains in part why the AUC-ROC values achieved in our study are low, although still better than chance. Interestingly the textural features, which describe the heterogeneity of DCIS lesions, had better predictive power when analyzed in conjunction with the shape or topological features. Tumoral heterogeneity at the genetic level has been described as being associated with invasive potential [39], but the relationship between imaging features and genetics, termed radiogenomics, has not been adequately explored. Further investigation into the relationship between genetic heterogeneity and phenotypic manifestations in the form of calcifications at mammography are needed. These results indicate that computer-vision extracted mammographic features associated with microcalcifications and clusters show potential to predict invasive disease and if features can be designed specifically for the task of predicting upstaging the performance metrics will also likely improve.

The physician-interpreted features in our study showed mixed results for the prediction of upstaging. Common features such as BI-RADS morphology and distribution, which are associated with the risk of malignancy, were not useful in stratifying the risk of occult invasive disease. The pathology features including nuclear grade and presence of comedo type necrosis were also not associated with upstaging, despite a prior meta-analysis by Brennan et al. [5] which demonstrated an association with invasive cancer; however, this may be due in part to the smaller sample size included in our study. Previous investigators have looked at the predictive power of individual histologic or pathologic features [5, 6, 9, 25, 27–29], but only Park et al. [9] constructed a nomogram to predict the occult invasiveness of DCIS patients, and obtained an AUC-ROC of 0.71 on their validation set. Although similar in performance to our current study, the previous study included DCIS patients with microinvasion, and the overall underestimation rate of DCIS patients by preoperative biopsy was 42.6% (145/340), both of which significantly decreased the difficulty of the prediction task. A unique feature of our study is the inclusion of the radiologist’s subjective score of being invasive, which did show an association with upstaging for one of the two radiologists. This suggests that the radiologist might be able to identify predictive features that are not expressed in the BI-RADS descriptors, and provides an opportunity for computer-vision features to attempt to quantify and potentially improve upon what the radiologist is interpreting. This is especially important because when included into a multivariate model, the physician-interpreted features did not perform better than random chance.

The current study has several limitations. First, the sample size is relatively small because of the single-institution design and the strict exclusion criteria. This affects the performance of our predictive model, especially considering the high-dimensional computer-vision feature set we extracted. Due to this, a robust and effective feature selection method is necessary. We investigated several different methods and found the sequential floating forward feature selection method with the area under the precision-recall curve as the optimization measure [40] provided the best result (the range of AUC-ROC for other tested feature selection methods are [0.55 ~0.7]). Second, we did not conduct a thorough inter-observer analysis for radiologist-interpreted features, because the main purpose of this study was to explore the feasibility of building an automated model based on computer-extracted mammographic features. However, since the current two radiologists showed interestingly different performance in predicting occult invasive disease in DCIS, a comprehensive inter-observer analysis that includes more radiologists is warranted. Finally, this preliminary study was an unblinded, retrospective study. We were intentionally conservative with feature selection and modeling algorithms, as well as limiting the number of trials. However, our reported results were not adjusted by a multiple-hypothesis testing correction procedure, so there is still potential for overfitting bias.

Conclusions

In conclusion, our study demonstrates the feasibility of using computer-vision algorithms to predict occult invasive disease in DCIS. As a preliminary investigation, this work directly utilized CAD techniques and features not specifically designed for this task, and the size of the data set used is also relatively small, thus we were intentionally conservative with feature selection and modeling so as to minimize overfitting bias. In spite of these constraints, we were able to statistically predict occult invasive disease better than chance. Future work includes building upon these promising results in order to both refine and design new computer vision features specifically targeted towards detecting occult invasive disease, with the goal of improving the predictive power of our algorithm. In addition, we are also planning to include more subjects into this study via collaboration with other institutions for a better-justified validation of our model. Since the successful prediction of occult invasive disease in patients with newly diagnosed DCIS could notably change clinical management, therapeutic decision making, and discussions regarding patient outcomes, there is great potential for this line of inquiry.