Table 5.

Comparison of proposed approach with state-of-the art methods applied for HER2 scoring

Ref.	Dataset used	Methods/features	Classifiers	Accuracy
Pitkäaho, Lehtimäki et al.	119 core regions extracted from 81 WSIs	Data augmentation, block-based Scoring by CNN	CNN, AlexNet architecture	Accuracy 97.7%
Singh and Mukundan	1345 image patches from 52 WSIs	Intensity and colour features ULBP	Neural network classifier	Accuracy 91.1%
Cordeiro, Ioshii et al.	2580 patches from 86 WSIs	Image patch level and patient level scoring with colour and texture features	SVM, KNN, MLP and decision tree were compared	Best Accuracy 94.2%
Saha and Chakraborty	752 images cropped from 79 WSIs	Image patch-based nuclei detection and cell membrane extraction	Her2net—LSTM recurrent network	Segmentation and classification Accuracy 98.33%
Mukundan	4019 image patches from 52 WSIs	Characteristic curves, ULBP connectedness, entropy	Logistic regression, SVM	Average 91% Maximum 93%
Khameneh, Razavi et al.	127 WSIs	Super-pixel for tissue region extraction Colour and texture features	Modified Unet architecture for tissue classification	Classification Accuracy 87%
Proposed approach—transfer learning and statistical voting	2130 training patches and 800 test patches from 40 WSIs	Image patch-based labelling using transfer learning followed by statistical mode for scoring	VGG19 Architecture followed by fully connected dense layers for 3 class	Test data of 100 images Patch-based accuracy 93% Image-based accuracy 98%