Resiniferatoxin br Tumour normal segmentation br Fig illustr
4.4. Tumour/normal segmentation
Fig. 8 illustrates the ability of our approach to segment tumour epithelium in WSIs that Resiniferatoxin contain tumour epithelium only (a,d), and both normal and tumour epithelial regions (b,c). We used a sin-gle training image here, that of Fig. 7(a). By including more im-ages in the training set as described in Section 4.2, the accuracy is slightly increased resulting in better overall performance, as shown in Table 2.
Table 3 compares the performance of all 6 approaches on the 8 WSIs of Fig. 7 (f–m). Note how our algorithm outperforms the others in terms of precision, recall, and dice metrics.
We also compared all 6 approaches on the TMA set. Our ap-proach outperformed the others in terms of median, as shown in Fig. 9. A One-way, repeated measures ANOVA showed that the methods were statistically significantly different (p < 0.001) with planned comparisons indicating that our approach was statisti-cally significantly better than the others (all Tukey corrected p − values < 0.01).
To illustrate the effect of FSPF, and our proposed WAI-based and ALI-based features, Fig. 10 shows the segmentation result of a nor-mal TMA image, where the normal region is highlighted in red while the misclassified region is highlighted in green, with differ-ent versions of the proposed approach to test each of its compo-nents independently. Note how the performance of our approach is significantly improved by combining both ALI-based and WAI-based features. The same conclusions can be confirmed when we compare the accuracy of these modified versions of the proposed approach, in terms of precision, recall, and dice metrics, as illus-trated by Table 4.
The Precision (%), Recall (%), and Dice (%) metrics for different versions of our pro-posed approach applied to a TMA image of Fig. 10.
The robustness of the proposed approach to noise: the precision, recall, and dice metrics obtained when different Gaussian noise levels controlled by standard devi-ation (SD) and salt & pepper noise model.
To illustrate the robustness of the SOM-based classifier to noise, we compared the behaviour of our approach to a similar one with a REF − SV M classifier instead. This is by first corrupting a test WSI with different noise models, and then using the trained RBF − SV M and SOM-based classifiers to segment out tumour ep-ithelium of a testing image. Fig. 11 illustrates the performance of our approach on a single WSI corrupted with Salt & Pepper noise model (e,j), and increasing levels of Gaussian noise: SD = 5 (a,f) to SD = 35 (d,i). Note receptor our approach is more robust to noise when using SOMs as a classifier than when using a RBF − SV M classifier, as confirmed in Table 5.
To illustrate the impact of noise at different features, we cor-rupted the individual features using white Gaussian noise, in both training and testing phases, and evaluated the classification perfor-mance of our proposed approach on a WSI of Fig. 11(a). Tables 6 and 7 show the effect of noise at each corrupted feature in terms of precision, recall, and dice metrics in the testing and training phases, respectively. Note how robust our proposed ap-proach in dealing with attribute/feature noise. As demonstrated by Tables 6 and 7, we can see that attribute noise in the training phase has a higher impact on the classification performance than in the testing phase. Also ALI and WAI5 are the most sensitive features to noise.
Despite the fact that SOM-based classifier has the ability to ef-fectively preserve the topological structure of the training data us-ing a few sets of prototypes and hence its robustness to noise. However, similar to other prototype-based classifiers, our ap-proach is still sensitive to the very high-level of noise as shown in Tables 5 and 6. Interested readers may refer to Zhu and Wu (2004) for the different possible solutions to deal with noise, especially to detect and correct attribute noise.