Automatic evaluation of tumor budding in immunohistochemically stained colorectal carcinomas and correlation to clinical outcome

Abstract

Background: Tumor budding, meaning a detachment of tumor cells at the invasion front of colorectal carcinoma (CRC) into single cells or clusters (<=5 tumor cells), has been shown to correlate to an inferior clinical outcome by several independent studies. Therefore, it has been discussed as a complementary prognostic factor to the TNM staging system, and it is already included in national guidelines as an additional prognostic parameter. However, its application by manual evaluation in routine pathology is hampered due to the use of several slightly different assessment systems, a time-consuming manual counting process and a high inter-observer variability. Hence, we established and validated an automatic image processing approach to reliably quantify tumor budding in immunohistochemically (IHC) stained sections of CRC samples.

Methods: This approach combines classical segmentation methods (like morphological operations) and machine learning techniques (k-means and hierarchical clustering, convolutional neural networks) to reliably detect tumor buds in colorectal carcinoma samples immunohistochemically stained for pan-cytokeratin. As a possible application, we tested it on whole-slide images as well as on tissue microarrays (TMA) from a clinically well-annotated CRC cohort.

Results: Our automatic tumor budding evaluation tool detected the absolute number of tumor buds per image with a very good correlation to the manually segmented ground truth (R2 value of 0.86). Furthermore the automatic evaluation of whole-slide images from 20 CRC-patients, we found that neither the detected number of tumor buds at the invasion front nor the number in hotspots was associated with the nodal status. However, the number of spatial clusters of tumor buds (budding hotspots) significantly correlated to the nodal status (p-value = 0.003 for N0 vs. N1/N2). TMAs were not feasible for tumor budding evaluation, as the spatial relationship of tumor buds (especially hotspots) was not preserved.

Conclusions: Automatic image processing is a feasible and valid assessment tool for tumor budding in CRC on whole-slide images. Interestingly, only the spatial clustering of the tumor buds in hotspots (and especially the number of hotspots) and not the absolute number of tumor buds showed a clinically relevant correlation with patient outcome in our data.

Keywords: Colorectal carcinoma; Convolutional neural network; Digital pathology; Image processing; Tumor budding.

Fig. 1

Correlation of the budding score to clinical data. The externally provided budding score of 328 cases showed a correlation to the patient’s nodal status (a) and to the regression-free survival (b), as previously published by [9, 10]. Furthermore, within the database there is also a significant correlation between the budding score (grades 1 to 4 with the intervals [0–5], [5–17], [17–20] and [20-∞] to stratify the absolute number of tumor buds [9]) and the morphology-based tumor grading (Grade 1–4) (c)

Fig. 2

Definition and detection of tumor buds. a Example single TMA core stained for pan-cytokeratin and hematoxylin and eosin (HE). The red circle highlights a tumor bud next to larger tumor island. The blue circle is a small tumor part, which we referred to as a “tumor islet”; which is a small tumor aggregate but does not meet all criteria for a tumor bud. b Histogram of pan-cytokeratin positive tumor area per core (upper plot) and pan-cytokeratin tumor bud area per core (lower plot). One can see that the mean area of all tumor fragments (upper panel) shows a huge dispersion (x-axis 0–100,000), whereas the mean area of the tumor buds is less distributed (x-axis 0–1000)

Fig. 3

Sketch of the analysis. A) ROI tumor (not shown) and the ROI border (dashed red line) were manually delineated by a pathologist (CW). A1: a sliding window moved over the ROI border and cropped every half TMA diameter in the underlying, TMA core-sized image. A2: The resulting tile (exactly the size of one TMA core) was then defined as a virtual TMA (vTMA). b, c This procedure leads to a value (number of tumor buds or budding score) per vTMA. This spatial data could be plotted as a heat map (example heat maps with arbitrary values in the left part) or as a histogram (right part of the figure)

Fig. 4

Analysis of vTMAs from 20 cases. From 20 selected cases (pN0 (n = 9), pN1 (n = 5), pN2 (n = 6)) one pan-cytokeratin-stained slide was digitalized and disassembled into virtual, overlapping vTMA cores (n = 290 ± 152). a No significant correlation was detected between the resulting budding score and the nodal status and b the median number of tumor buds within the 10 hottest spots and the nodal status for the complete ROI border. c, d Significant positive correlation between the nodal status and the absolute number of significant budding hotspots and the normalized number of significant budding hotspots. The latter is done to compensate for a trend toward higher tumor areas on the WSI for pN1 and pN2 cases

Fig. 5

Monte Carlo-like simulation. From every case (of the 20 cases with 290 ± 152 vTMAs from the tumor border) repetitively a predefined number of vTMAs were chosen at random, and subsequently the median number of buds, the median budding score and the normalized entropy were calculated. This process was repeated several times (n = 10) with different sample sizes (from n = 2 to n = 200). a and c show the normalized entropy and the median number of tumor buds in relation to the sample size, respectively. b and d show the normalized entropy and the median number of tumor buds for the complete tumor border, respectively

Fig. 6

Correlation of the TMA data to clinical data. a Histogram of the number of TMA cores from the tumor region per case (n = 4 ± 2). b, c Histogram of the b median budding score per case [0–4] and the c maximal budding score per case. The nodal status [pN0–2] is added by color coding