The Histological Diagnosis of Breast Cancer by Employing scale invariant ResNet 18 With Spatial Supervised Technique

Abstract Background Breast cancer is one of the most prevalent cause of morbidity and mortality in women all over the world. Histopathological diagnosis is a vital component in the management of breast cancer. The application of artificial intelligence is yielding promising results for the better patient care. Aim The main aim of the present research project is to explore the potential of spatial supervised technique to develop scale invariant system for the histological diagnosis of breast cancer. Materials and Methods The anonymized images of hematoxylin and eosin stained section of the dataset, which has been acquired from the website. The slides were taken at different zoom (magnification) levels. Spatial supervised learning has been employed to make a scale invariant system. We used 400x and 40x to generate the results. For the 400x, we trained our network on a dataset of 200x,100x, and 40x images. The datasets were split into training and validation sets. The training set contained 80% digital slides of the respected dataset, and the validation set contained 20% digital slides of the respected dataset. The final result was generated by splitting the dataset of 400x into the training and test dataset. The training set contained 50% digital slides, and the test set also contained 50% digital slides. This unusual split is done to show how good spatial supervised learning works. Similarly, for 40x, we trained our networks on a dataset of 400x,200x, and 100x. The same steps were followed to obtain the 40x results. Results The result analysis revealed that the ResNet 18 with spatial supervised learning on dataset of 40x yielded the F-1 score of 1.0, while ResNet 18 with supervised learning only, on dataset of 40x yielded F-1 score of 0.9823. ResNet 18 with spatial supervised learning on dataset of 400x revealed F-1 score of 0.9957, and ResNet 18 with supervised learning only, on dataset of 400x showed the F-1 score of 0.9591. For supervised learning dataset is spited into training (80%) and testing (20% of dataset). Conclusion The analysis of digitized pathology images with the application of convolutional neural network Resnet -18 architecture with spatial supervised learning revealed excellent results, which is demonstrated by a very high F-1 score of 1.0. The development of scale invariant system with application of spatial supervised technique solved the problem of images with variable magnifications. The finding would further pave the pathway for application of deep learning for the histological diagnosis of pathological lesions.