br tained using a Hamamatsu NanoZoomer with
tained using a Hamamatsu NanoZoomer with 40X magnifica-
tion (0.5 μm resolution) and downsampled to approximately
All sections were obtained from previous clinical studies. All tis-
(20) sue sections underwent immunohistochemical staining using dif-
ferent protocols (optimized for each specific antibody) but all
used3’3-Diaminobenzidine as a chromogen and all were coun-
terstained with Mayer’s Haematoxylin. The staining pattern of
the Filipin III was as follows: CD3- membranous staining of T-
(21) lymphocytes, BerEP4- cytoplasmic staining of colonic epithelium
(both tumour and non-tumour), AE1/3- cytoplasmic staining of
colonic epithelium (both tumour and non-tumour), MLH1 - strong
staining of tumour nuclei and weak staining of nuclei of normal
epithelium and stroma, MSH2 - strong staining of tumour nuclei
and weak staining of nuclei of normal epithelium and stroma. The
large array of antibodies with the varying staining patterns and
the combination of TMA/WSI were chosen in order to ensure that
the algorithms developed were not influenced by the immunohis-
tochemical test being performed or the nature of the tissue section.
Access to tissues and ethics approval were granted by Nottingham
Health Sciences Biobank which has approval as an IRB from North
West - Greater Manchester Central Research Ethics CommitteeREC
In order to compare the performance of our approach against
that of competing methods from the literature, we implemented
the following algorithms in Matlab on the basis of the infor-
(22) mation from
the associated papers: the CMF-based algorithm
based and Higher-order Histogram (HH)-based algorithms of
Kather et al. (2016). We also ran the publicly available code of
the following methods: the Local binary patterns (LBP)-based al-
gorithm of Linder et al. (2012), the Perception-based algorithm of
We adopted Precision, Recall, and Dice similarity coe cient
metrics calculated at the pixel level to quantitatively assess the
TP + FN
where TP, FP, and FN represent the numbers of true positive, false positive, and false negative tumour epithelium or normal epithe-lium pixels. Dice coe cient measures the spatial overlap between machine vision result and ground truth and hence quantifies the overall accuracy. Precision and Recall are sensitive to the amount of over-segmentation and under-segmentation, respectively, in the sense that over-segmentation (of the epithelial area and/or tumour epithelium) is associated with a small Precision score, whereas under-segmentation leads to a small Recall score.
Fig. 7. The WSIs used in the training (a-e) and testing (f-m) phases by our approach and other methods.
All experiments were run on a PC with the following configura-tion: 3.5 GHz Intel(R) Xeon(R), and 32.0 GB RAM in Matlab R2016b. Ground truth was obtained from the images by manual segmenta-tion: a trained operator separated the epithelium from the stroma and discriminated between normal epithelium and tumour epithe-lium. The developed code and dataset will be publicly available at http://www.aidpath.eu.
4.2. Parameter settings and training
Our dataset has been divided into training and testing datasets (respectively 10% and 90% of the total dataset). For example, a set of 5 WSIs (Fig. 7) each contains both normal and tumour regions have been selected for training our approach in addition to a set of 8 randomly selected TMAs images (4 of which are for normal cases). On the other hand, a set 8 WSIs and 126 TMA images used as a testing set for segmentation performance evaluation. We used the remaining images to train our model and fine-tune its param-eters. Then the parameters were fixed for all testing images as fol-low: local Gaussian intensity of FSPF (σ ) was fixed to 0.1, con-tour smoothing parameter(σ ) was fixed to 1.0, weight parameters (i.e., λ+ , λ− ) were fixed to 1.0, local Gaussian intensity of WAI (σ ) was fixed to 2, number of neurons in the output layer was fixed to 3 × 3, and starting learning rate (η(0)) was fixed to 0.1.