| Summary: | Introduction/ Background
The quantification of immunohistochemical Ki67 and the detection of active areas of tumor cell proliferation (hotspots) have a critical importance in the prognosis and treatment planning for breast cancer.
Aims
In this work an automated and robust method for the detection of hotspot areas in breast cancer biopsies is proposed with the aim of supporting the pathologists by highlighting hotspot areas.
Methods
The proposed method has been tested on one Ki67 stained image from Openslide [1] acquired at 40x with an Hamamatsu scanner and on 5 Ki67 stained images of breast cancer biopsies acquired at 40x with AperioCS. Each input image is divided in tiles, whose colors are deconvolved using the method of Ruifrok [2] to estimate the presence of Ki67. Tiles with an estimated Diaminobenzidine (DAB) positivity of more than 5% are considered as potential hotspots. Neighbouring positive tiles are merged to form a final hotspot area. The three hotspot areas with higher DAB positivity are also highlighted in the output. The hotspot areas are then written in an XML file which is read by a medical image viewer such as Aperio ImageScope. The proposed method has been implemented in Java using the BioFormats open source library [3].
Results
The color deconvolution of each tile has been performed by applying the standard Hematoxylin/Diaminobenzidine (H/DAB) deconvolution matrix provided in Fiji [4]. The tests have been performed at different zoom levels resulting in similar, coherent outputs of hotspot areas. The resulting hotspot areas have been validated by visual comparison with the hotspot areas determined by experts, showing a significant superimposition as shown in Fig. 1. The proposed method has been compared with the ASH method [5] for the image acquired with Hamamatsu scanner, producing similar results in comparable time.
In conclusion, a new fully automated method for the detection of Ki67 hotspot areas in breast cancer biopsies has been proposed to support the pathologist by highlighting different hotspot areas. It is able to process different medical images formats, making it more interoperable. The time performance is comparable with existing methods such as ASH [5]. The proposed method at low magnification such as 1x, 2x, and 4x produced varying results influenced by the size that each tile assumes at such magnification levels. Further extension of the proposed method will also include the MIB-1 estimation within the hotspot areas as well as extensive testing and validation. This work is partially founded by the EU FP7 program, grant number 612471.
|
|---|
