Women's College Hospital began the routine use of mammography in the mid-1960's, establishing a breast imaging expertise. This led to increased detection of more breast DCIS at the Henrietta Banting Breast Cancer Center, a multidisciplinary assessment center for breast diseases, in the study period between 1979 and 1994 . Analysis of patient records from the practices of the group of teaching surgeons identified 260 women who were diagnosed as having DCIS of the breast. Study cases included had (a) DCIS confirmed by pathology review, (b) histology slides of the initial DCIS and most subsequent carcinomas available for review, (c) no previous breast or other malignancy, and (d) a detailed follow-up to 1997 . One hundred and eighteen patients were excluded for the following reasons: 18 on review did not have DCIS; 100 had previous carcinoma, and 18 had no (or limited) follow-up or primary histologic slides were not available for review. The data for the remaining 124 patients with DCIS formed the basis of the previous study. These patients had a median 5.0 years of follow-up. The focus was the 88 patients who underwent lumpectomy alone since this group experienced most of the subsequent clinical events: 17 of 19 recurrences of DCIS, with the other 2 observed in 18 patients who underwent lumpectomy followed by adjuvant radiotherapy, and all 19 DCIS recurrences were ipsilateral at median interval 2.6 years, range 5.3 months to 5.7 years; 11 of 12 developments of invasive carcinoma, the eleven invasive cancers were 6 ipsilateral and 5 contralateral at median 1.8 years, range 9.5 months to 6.6 years. The median time for development of ipsilateral invasive carcinoma was 1.6 years, range 9.5 months to 6.5 years. The twelfth diagnosis of invasive carcinoma in the original DCIS series was an ipsilateral axillary lymph node, which occurred in a patient who was initially treated with simple mastectomy  and who is excluded from these investigations. Seventy-eight of the 88 patients (88%) were detected mammographically. None of the 88 patients received adjuvant radiotherapy or systemic therapy.
The specimens were processed uniformly in a manner consistent with standards at the time of the biopsies . Specimens were fixed in 10% neutral buffered formalin. Tissue blocks were created with uniform section thickness of 3–4 microns. Tissue evaluated had not been examined at previous frozen section. For mammographically detected lesions, tissue was sampled rather than assessed in toto, with sampling directed to the area marked by dye instilled preoperatively and/or area marked by a needle placed intraoperatively by a radiologist. Sampling of other specimens was directed by the gross appearance of the specimen. Several assessments were made to reflect size: (a) estimate from the gross description, (b) maximum dimension per slide, and (c) number of slides with DCIS involvement.
The percent parenchyma involved (<10%, 10 to 50%, >50%) was assessed to reflect the proportion of the total parenchyma (stroma and all ducts and lobules) in the areas on the slides containing DCIS that was occupied by the involved ducts. Percentage parenchyma was based on both fibrous and fatty stroma, and was determined on sections which contained DCIS. This reflected whether the involved ducts were concentrated within the parenchyma (higher percentage) or more diffusely scattered (lower percentage). Duct distribution tended to be uniform, and a single categorization could be assigned to each patient.
Results were obtained by considering the maximum size of the DCIS on any slide and the percentage of the parenchyma involved.
Resection margins had been painted with silver nitrate. The shortest distance between an involved duct and resection margin was measured microscopically with an ocular micrometer (in millimeters). The presence of an uninvolved duct between DCIS and painted margin (reported as cannot assess, not present, present) was also assessed.
The DCIS was classified into architectural patterns: solid, cribriform, micropapillary, other (e.g., papillary, apocrine, clinging). All of the types present, the predominant (most extensive) type present, and the type with least architectural differentiation (solid versus all other) were recorded. We used the results obtained by considering predominant architecture and least architectural differentiation: from previous paper , 51% of tumours were solid by worst architecture, while 33% were solid by predominant architecture.
Calcifications included present (amorphous or crystalline) or not present. Necrosis was central confluent (comedo) versus not: 67% of patients had necrosis.
DCIS was graded as 1, 2, and 3 (by the Van Nuys classification scheme . When more than one grade was present, the worst (highest) grade was recorded, as well as all grades present and the predominant (most extensive) grade.
Good quality computer images for the purpose of image analysis could be obtained from the archived hematoxylin and eosin (H&E) stained slides of 80 of these patients; 3 patients did not have slides available for assessment, while 5 had H&E stained slides which resulted in poor quality images that were unsuitable for further evaluation. These 80 patients are the subset considered for the current study. Digital images of the slides of the 80 study patients were acquired under the supervision of a breast pathologist (NM).
Representative H&E stained slides for each patient were selected that demonstrated the nuclear grading previously observed for that patient. A maximum of 2 slides per patient biopsy were selected. Two fields were selected so that ducts were sufficiently concentrated to contain a minimum of 5 ducts per field in which the nuclear grading was represented (either two fields in one slide, or one field on each of 2 slides). Affected duct spaces were contiguous. A low, 10 times, magnification was used for identification of appropriate sampling regions, while 40 times magnification was used for capturing images. A computer image was acquired for each of 5 ducts in one field and this was repeated for 5 ducts in the other field. Image features were measured for each of approximately 20 representative nuclei per duct, for a total of about 200 nuclei for each patient. Thirty-nine computer image features were extracted for each nucleus that described morphometry (size and shape of nuclei), densitometry (amount of staining of the nuclei), and texture (arrangement of staining in the nucleus). Images were acquired using NIH-Image software v.1.57 written by Wayne Rasband . Nuclear morphologic and densitometric features were measured with NIH-Image v.1.62b34-Arnv software. Texture features were measured using TextureCalc v.1.1ax software written by W.C.-C. All nuclear images were segmented by a single person (DA). Some image feature calculations and the merging of all nuclear image feature data to per duct, per field and per person attribution were accomplished with StatView v 5.01 (Brain Power, Calabasas, CA) software. Nuclear images were segmented for image analysis without knowledge of the corresponding clinical features or pathologist's grading. Nuclei distributed throughout the image field were segmented in order to assure a representative sample. Nuclear images that were incomplete (cut off at the edges of the field) or overlapping were not included. The few image fields that were indistinct or out of focus were not used.
Image analysis features
For each nucleus, 39 features were determined in three categories. (i) Morphometry: area, perimeter, ellipse major axis, ellipse minor axis, ellipticity (major axis/minor axis), shape form factor (4 × pi × area/perimeter squared), and roundness b (4 × area/pi × ellipticity squared) . (ii) Densitometry: mean density, standard deviation of density, modal density, minimum density, maximum density, sum density (mean density × area, used instead of I.O.D. of NIH-Image), range density. (iii) Markovian texture features [17, 18] were calculated from the Markovian co-occurrence matrix of pixel densities with a step size of 2. They were angular second moment, contrast, correlation, variance, inverse difference moment, sum average, sum variance (corrected from ), difference average, difference variance, initial entropy, final entropy, entropy, sum entropy, difference entropy, coefficient of variation, peak transition probability, diagonal variance, diagonal moment, second diagonal moment, product moment, and triangular symmetry. Additional texture features, calculated from the binned histogram of pixel gray scale values, included histogram mean, histogram variance, histogram skewness, and histogram kurtosis. The mathematical formulae defining these image features can be found in the cited references.
The clinical factors recorded on these patients were age (in years) and type of presentation (mammographic, clinically palpable, bloody nipple discharge). The histologic factors previously evaluated  were maximum DCIS size (cm), percentage of parenchyma involved with DCIS (<10%, 10–50%, >50%), predominant architecture (0 – cribriform/micropapillary/other, 1 – solid), worst architecture (0 – cribriform/micropapillary/other, 1 – solid), nuclear grade [by the Van Nuys Classification system worst (nuclear grade 1, 2, 3); also, predominant (nuclear grade 1, 2, 3)], necrosis [none, confluent (comedo-like)], calcification (none, crystalline/amorphous), measured margin (zero margin, <1 mm, 1–5 mm, >5 mm), presence of uninvolved intervening duct (not assessable, no, yes), Van Nuys Prognostic Index. In addition, 39 nuclear image features were determined for about 200 nuclei per patient.
For each patient, the image data were pooled across i) all nuclei in a duct (10 assessments), ii) all nuclei in a field (2 assessments), and iii) all nuclei for a patient (1 assessment) to yield a summary feature value [adjusted mean = mean/(standard error of the mean)], for each of the 39 image features for nuclei of the 13 different assessments per patient: 10 ducts, 2 fields, 1 overall. In addition, grading discriminant classification functions, that are weighted combinations of image features, described below in the Analysis section, were assessed as prognostic factors.
Thirteen different assessments, corresponding to the 13 different ways of pooling the image analysis feature data, were performed to examine the effects of DCIS heterogeneity on apparent associations with clinical outcome. In other contexts, investigations have been restricted to single ducts, fields, or pooled per person assessments without an examination of replicability.
A new diagnosis of breast carcinoma made more than 90 days after the initial surgery was designated as an event. Invasive carcinoma in this group of patients occurred about equally in both the ipsilateral and contralateral breast which is consistent with the findings of some others [19–21], including those for potentially lower risk DCIS patients . Using t-tests, there was no statistically significant difference in image analysis features between patients who developed invasive disease ipsilaterally, as opposed to contralaterally. For these investigations, an event was considered to be development of invasive carcinoma whether ipsilateral or contralateral. There were no deaths from breast cancer, or another cause, in this group of patients over the study period.
Statistical analyses were performed with BMDP PC Dynamic Version 7.0 (same as BMDP-XP, Statistical Solutions, Sagua, MA).
Image analysis pre-processing of data included for each image feature and each patient, 1) Levene's tests for equality of variance between ducts and fields for each person and between people, 2) the use of the mean/S.E.M. of image features on a duct, field and patient basis, resulting from indications in Levene's tests of significant evidence against assumption of equal variances, 3) per duct, per field, and per person grading disciminant classification functions from forward step-wise Fisher linear discriminant analyses, using an entry p-value of p ≤ 0.05, and 4) assessment of the correct classification by the grading discriminant classification functions using jackknifed (leave-one-out) classification of patients. Standardized coefficients for canonical variables in the discriminant function are reported.
The histologic, clinical, and image analysis factors were assessed with respect to whether they were associated with the development of invasive disease. Univariate assessments were with Kaplan-Meier plots and the Wilcoxon (Peto-Prentice) test statistic. For each image feature, standard image analysis cut-points at the means of the data were utilized after confirmation that the data were approximately symmetric.
Multivariate assessments were with Cox forward step-wise regressions, using the likelihood ratio criterion (~χ2
(1), p ≤ 0.05) as the test statistic to determine if a factor would be added to the model. Since we had no knowledge of which of the image analysis features assessed would best reflect a patient's DCIS, or the extent to which differences in image features might relate to prognoses, we performed 13 sets of multivariate analyses, corresponding to the 13 generations of image feature factors per patient: per 10 ducts, 2 fields, 1 pooled across 2 fields.
This study was approved by the Ethics Review Board at Women's College Hospital, Toronto, Ontario, Canada.