Assessment of Ki-67 proliferation index with deep learning in DCIS (ductal carcinoma in situ)

Abstract

The proliferation index (PI) is crucial in histopathologic diagnostics, in particular tumors. It is calculated based on Ki-67 protein expression by immunohistochemistry. PI is routinely evaluated by a visual assessment of the sample by a pathologist. However, this approach is far from ideal due to its poor intra- and interobserver variability and time-consuming. These factors force the community to seek out more precise solutions. Virtual pathology as being increasingly popular in diagnostics, armed with artificial intelligence, may potentially address this issue. The proposed solution calculates the Ki-67 proliferation index by utilizing a deep learning model and fuzzy-set interpretations for hot-spots detection. The obtained region-of-interest is then used to segment relevant cells via classical methods of image processing. The index value is approximated by relating the total surface area occupied by immunopositive cells to the total surface area of relevant cells. The achieved results are compared to the manual calculation of the Ki-67 index made by a domain expert. To increase results reliability, we trained several models in a threefold manner and compared the impact of different hyper-parameters. Our best-proposed method estimates PI with 0.024 mean absolute error, which gives a significant advantage over the current state-of-the-art solution.

Figure 1

The complete experiment, divided into 4 stages. The primary source of data is a collection of 95 whole slide images (WSI). For each case, one hot-spot was selected and designated as a rectangular area, i.e. pROI (pathological region of interest) by a domain expert. Then, the images are annotated by the same expert and split into 4 chunks. The First 3 chunks are used to train and evaluate models (in a k-fold manner). The last chunk is used only to evaluate. This means, that at the end, we had results from 45 images for a single model (20 images from testing chunk + 25 from the one chunk used as a testing subset) and 20 images for an ensemble model (only from the testing chunk because other images have been utilized by at least one componential model).

Figure 2

Example of IHC slide. In the red regions there are marked: (a) artifacts, (b) lymphocytes, (c) blood vessels, (d) collagen fibers and (e) tumour cells with high and low expression of Ki-67 protein (our assumed ROI).

Figure 3

The complete defuzzification process for (a) 48-, (b) 96-, (c) 192-pixels window size applied in the neural network stage.

Figure 4

The results of fragments classification: (a) F1-score medians and their distributions of compared methods evaluated on 20 images from the testing chunk, (b) AUROC (area under the receiver operating characteristics) of the component models (“single”) grouped by the window size.

Figure 5

The results of Ki-67 proliferation index estimations: (a) presents MAE medians and their distribution for a subset of 20 images with excluded outliers, (b) shows the impact of the applied biasing onto reducing outstanding Ki67 PI errors, (c) sets linear regression from the best methods (“single biased” and “ensemble unbiased”), (d) comparing our best solution (ensemble 96 px-window unbiased) with PathoNet and the “base” approach.

Figure 6

The method outputs overlaying the 3 source images: (a) is an input image (b) with domain expert’s annotations, (c) is a fuzzy output from the neural network stage, (d) is a binary mask being a result of the defuzzification process, (e) presents selected cells—immunopositive with a red outline and immunonegative with a green one, (f) PathoNet estimations as a reference. Our approach selects the nuclei area, while PathoNet the nuclei centers. It should be noted that our solution more accurately rejects lymphocytes and artifacts as non-tumor cells, and therefore, it approximates the proliferation index better. However, for our solution the most problematic are cells near the edges and staining blobs, which might bias the segmentation.