Nuclear grade, or all grades when there was more than one grade, was reported for each patient by the pathologist according to the Van Nuys system (Holland et al. 1994
; Silverstein et al. 1995
). This system has been compared to other systems that have been proposed to predict development of infiltrating carcinoma (Badve et al. 1998
; Leong et al. 2001
; Douglas-Jones et al. 1996
). Several of these grading systems have recently been reviewed (Leonard and Swain, 2004
). The greatest consistency among pathologists seems to be obtained with systems that are based, in large part, on nuclear grade (Morrow et al. 2002
) and a consensus conference on classification of DCIS recommended that DCIS should be stratified primarily by nuclear grade (Schwartz, 1997
). Our results indicate that image analysis can provide a reproducible quantitative description of nuclei for this purpose since duplicate measurements were similar for 89.7% to 97.4% of features assessed.
We investigated the ability of 39 image features describing nuclear morphology (size and shape), densitometry (amount of stain), and texture (arrangement of DNA) to quantitatively discriminate tumors with pathologically determined nuclear grade. The most representative grading for a patient utilized data obtained for nuclei in both fields. Features included in the discriminant function were one whose value was determined by the size and shape of the nuclei (minor ellipse axis) and one whose value was determined by the arrangement of DNA in the nuclei (peak transition probability). Similar rates of accurate classification of grade were obtained from the first field assessed (about 100 nuclei), the second field (about 100 nuclei), and both fields (about 200 nuclei), with correct classifications of respectively 65.0%, 67.1%, and 65.0% of the patients. However, the discriminant functions for the three situations differed by image features. Discrimination using both fields would be the most representative of the patient’s grading.
Discriminant analysis has frequently been used to classify patients based on feature values determined by computer-aided image cytometry (Baak et al. 1991
; Bartels, 1980
; Patterson, 1995
). The features selected and the weights assigned to each feature are readily interpretable. Further, we examined the relevance of the continuous discriminant functions here in a censored survival analysis framework.
In this study, image features were extracted from H&E stained slides rather than Feulgen stained slides. The Feulgen reaction is useful because it is specific for DNA, the intensity of the reaction is proportional to the amount of DNA, and nuclear regions of interest can be automatically segmented (Schulte et al. 1995
). However, conventional H&E stained slides have also been used to extract morphometric and other nuclear feature values (Frank et al. 2001
; Tuczek et al. 1996
; Wolberg et al. 1995
; Christen et al. 1993
; Peinta and Coffey, 1991
; Weyn et al. 2000
). The possibility of extracting nuclear features from H&E stained slides offers several advantages. First, existing H&E stained slides available from pathology archives can be used without recovering the original paraffin blocks, recutting tissue sections, and staining the new slides with the Feulgen procedure; second, digital images of each microscopic field can be acquired by the image analysis system simultaneously with review by a pathologist viewing familiar H&E stained tissue. Although the density measurements extracted from H&E stained slides are not directly proportional to amount of DNA as they are for Feulgen stain, we have found that nuclear area is proportional to amount of DNA as determined by Feulgen (r = 0.886). Nevertheless, measurement of nuclear area in H&E stained specimens is not an adequate substitute for measurement of DNA ploidy as determined with the DNA-specific Feulgen stain. In the future, the labor intensive chore of manually segmenting nuclear regions of H&E stained slides might be overcome by acquiring color images and segmenting them with appropriate algorithms (Ferr-Roca et al. 1998; Latson et al. 2003
). In our study only 5.7% of the archival slides had too poor stain quality for image analysis.
Patients with breast DCIS that had intermediate nuclear grades, and/or more than one grade, offer challenges to pathologists. Such patients may represent as much as 50% of DCIS patients (Miller et al. 2001
). In this study computer-aided image analysis was used to assess a cohort of patients who presented with DCIS, many with intermediate or mixed grades. H&E stained slides were viewed by a pathologist who recorded the nuclear grades in each field and digital images were acquired simultaneously. Nuclear image feature values were extracted and patients classified by discriminant analysis using the pathologist’s grouping of patients to supervise the analysis. Patients were placed into three groups according to the grades assigned by the pathologist taking into account the large proportion of patients with intermediate and/or mixed grades. The discriminant function correctly classify 78.1% of patients with low nuclear grade, and 70.6% of patients with high nuclear grade. The same discriminant function had a poorer success rate of only 48.4% for the intermediate group. This suggests that the intermediate group could be subtyped by characteristics different than those separating the low and high grade groups. Such subtypes of the intermediate grade group might be relevant for therapeutic decisions.
Since high nuclear grade is one of the factors that has been associated with recurrence it might be expected that high nuclear grade would also be associated with recurrence within this cohort of patients. However, neither the highest nuclear grade, nor the most prominent nuclear grade, was associated with recurrence (Miller et al. 2001
; Chapman et al. 2007
). The large proportion of patients with mixed nuclear grades may have contributed to the lack of association. In order to take into account the patients with mixed grades, discriminant analysis used the pathologist’s nuclear grade(s) for each patient, the grouping of patients with mixed grades, and the quantitative nuclear feature values determined by image cytometry. A discriminant function was derived that optimized the classification of patients into grading groups. The value of the canonical variable was determined for each patient. The canonical variable was standardized across patient values to have a mean of zero and a standard deviation of one. The distribution of patient values by canonical variable appeared to be continuous rather than three discrete and separate groups. The discriminant function was not associated with ipsilateral DCIS recurrence, although it had been with some pooled assessments concerning the development of invasive breast cancer (Chapman et al. 2007
). One of the image features included in the discriminant function was minor ellipse axis (p < 0.001); it is noteworthy that a related feature, ellipticity, was associated with DCIS recurrence (p = 0.04). Rounder nuclei were associated both with higher grade and DCIS recurrence. It should be noted that this association was found with image feature adjusted means, after accounting for the greater variability observed with high nuclear grade. The image features alone would not be expected to be sufficient for prognosis. However, the additional information, added to other pathologic and molecular features, may improve the prognostic classification of patients.