The incidence and mortality rates of cervical cancer have both significantly decreased as a result of routine screening. Retrospective review of cytology and HPV test results with cervical biopsy diagnosis is essential for validating and adapting these algorithms to changing technologies, demographics, and optimal clinical practices because selection of appropriate screening modalities depends on well-validated clinical decision algorithms. Due to the overwhelming number of specimens, however, manual categorization of the free-text biopsy diagnosis into discrete categories is extremely laborious, which may result in significant error and bias. Computer-based classification tasks have seen significant progress thanks to advances in natural language processing and machine learning, particularly in the last ten years. For the purpose of developing a supervised classifier that is capable of assigning precise biopsy categories to free-text biopsy interpretations while maintaining high concordance with manually annotated data, we use an effective version of an NLP framework called FastTextTM and an annotated cervical biopsy dataset to accomplish this. After a referee review by an experienced pathologist, we examine cases in which the machine-learning classifier disagrees with previous annotations. With a concordance of 97.7%, we also demonstrate that the classifier is robust on an untrained external dataset. In conclusion, we discuss the advantages and drawbacks of this strategy and demonstrate how NLP can be applied to a real-world task of pathology classification. In the field of digital pathology, transfer learning has been the most common method due to a lack of annotated pathological images. Pre-trained neural networks based on the ImageNet database are frequently used to extract "off-the-shelf" features, predicting tissue types, molecular characteristics, clinical outcomes, and other outcomes with great success. We conjecture that calibrating the pre-prepared models utilizing histopathological pictures could additionally further develop highlight extraction, and downstream forecast execution. In a two-step process, we fine-tuned a pre-trained Xception model using one million annotated H&E image patches for colorectal cancer (CRC). The Image-pretrained model and the finely tuned Xception model's features were compared using the following methods: 1) tissue classification using the same image type used for fine-tuning for H&E images from CRC; ( 2) expectation of invulnerable related quality articulation, and for lung adenocarcinoma. The model's performance was evaluated using five-fold cross validation. 50 times were spent on each experiment. For both cross-cancer prediction for gene expression and mutation at the patient level and for same-cancer tissue classification, where similar images from the same cancer are used for fine-tuning, we demonstrated that fine-tuning the pretrained ImageNet neural networks with histopathology images can produce higher quality features and better prediction performance.

Home page link: https://obstetrics.imedpub.com/

With Regards,
Sara Giselle
Associate Managing Editor
Journal of Critical Care Obsestrics & Gynocology