Article Preview
Top1 Introduction
Breast cancer is the most common malignant tumor in women and the most deadly. Breast cancer is a major public health problem in both industrialized and developing countries where its rate is increasingly growing year after year according to the 2020 statistics of the international agency for research on cancer1. Its diagnosis is always based on the histo-radio-clinical tripod and its management must absolutely be multidisciplinary.
Among the many different histological types that breast cancer involves, ductal and lobular carcinomas are distinguished and often oppose each other on different pathological criteria. Ductal and lobular carcinomas are the two most common histological types of invasive breast cancer. The ductal form dominates all histological types and accounts for more than half of all invasive carcinomas. The lobular form, second in frequency, accounts for only 5 to 10% of cancers.
Immunohistochemistry provided an appreciable comfort both to confirm the diagnosis of malignancy and to evaluate the loss of expression for E-cadherin for lobular carcinomas and positivity for other forms, notably ductal carcinomas.
However, traditional manual diagnosis requires an intense workload by competent experts. Misdiagnosis is likely to occur with pathologists who do not have sufficient diagnostic experience.
In this case, the need to evolve towards the use of computer-assisted-diagnosis (CAD) for the automatic classification of ductal and lobular cancers in order to improve the efficiency of the diagnosis will be noticed, but also physicians will be provided with results of more objective and accurate diagnostic results.
Deep learning techniques, such as convolutional neural networks (CNN), have shown great success in the detection of mitotic cells from histopathological images stained with hematoxylin and eosin (Ben Cheikh et al., 2017). This is a valuable source of inspiration for the development of an algorithm for classifying cancerous tissues from histopathological images colored with “E-cadherin”.
The objective of this work is to create a classification model dedicated to anapathic slides of cancerous breast tissue treated with E-cadherin antibodies to classify ductal and lobular carcinomas to help the pathologist to specify the sensitivity of cancerous tissues to the antibodies used and on the other hand to facilitate the work of oncologists for appropriate management and monitoring of patients
The adopted approach to solve this problem is based on transfer learning, which is a deep learning technique that a model trained for one task is reused for another related task in.
While it is demonstrated that all models are working well in practice, it is not clear that they work well when modified and used to accommodate small datasets. There has been considerable interest in this issue.
Small-SE-ResNet (Yun et al., 2019) is the basic model for this work. It is a very deep learning model trained on a large dataset (BreakHis) (Fabio et al., 2016) and it gives a very good performance both on the accuracy and the reduced number of parameters.
The proposed approach is to adapt it with a small dataset based on simple modifications but appropriate for better performance. Identical to (Yun et al., 2019), the increase in data was used to expand the number of images to improve the learning operation.
Our work is more original compared to previous related works in terms of time and economy. Existing works try to help doctors either by automatic classification or by detection and segmentation, which can give a clearer view of the slide. The novelty of our work is to make an immunohistochemical classification without using antibodies. Which, of course, will allow us to save the cost of these antibodies and at the same time save the time of fixation that will in return influence the speed of diagnosis.
To clarify this study, the paper has been organized as follows:
After the introduction an overview was given on image classification, deep learning, transfer learning and some related work. Then methods and materials section describe the objectives and design of the study, also the dataset used in this work and how to increase it with data augmentation were presented. Afterwards the Section 5 shows the proposed transfer learning methods. The article ends with a discussion and validation of the results. Finally, Section 6 concludes this study.