Advances in Applied Science Research Open Access

Trends in Medical Image Classification and Semi-Supervised Learning

Tian Guo Zhong^* Department of Physics, Jinan University, China
^*Correspondence: Tian Guo Zhong, Department of Physics, Jinan University, China, Email:

Received: 01-Nov-2022, Manuscript No. AASRFC-22-15202; Editor assigned: 03-Nov-2022, Pre QC No. AASRFC-22-15202 (PQ); Reviewed: 17-Nov-2023, QC No. AASRFC-22-15202; Revised: 22-Nov-2022, Manuscript No. AASRFC-22-15202 (R); Published: 29-Nov-2022, DOI: 10.36648/0976-8610.13.11.100

INTRODUCTION

Training machine learning and deep learning models for medical image classification is difficult due to a shortage of big, high-quality labelled datasets. Because medical experts spend a significant amount of time and effort categorizing medical pictures, models must be designed to train on relatively little labelled data. As a result, one viable option is to use Semi-Supervised Learning (SSL) methods. Successful predictions are obtained by mixing a small number of labelled datasets with a significantly larger number of unlabeled datasets. This enables SSL approaches to employ unsupervised learning knowledge to improve the supervised model. This research examines in depth the most recent SSL techniques offered for medical picture classification challenges.

The application of machine learning and deep learning algorithms in different fields of medicine, such as diagnosis, outcome prediction, and clinical decision-making, has been enabled in recent years by an increase in digitalization of health information and a progressive increase in processing capacity. Medical image analysis is a hot topic of machine learning and deep learning research since medical images are an important part of patients’ electronic health records. Currently, radiologists and other healthcare professionals manually extract information from medical images, making the procedure time-consuming and reliant on the individual’s experience and knowledge.

There has been a lot of research on ways to apply machine learning and deep learning in medical image analysis to produce better, faster, and more precise findings to solve these difficulties. In the area of medical image analysis, tasks such as segmentation, detection, de-noising, reconstruction, and classification are being addressed. This survey paper focuses solely on the most recent picture classification research.

Description

The models behaved differently with differing quantities of labelled data, as expected. For example, SRC-MT and NoTeacher demonstrated a score difference of up to 10% between 1% and 10% of labelled data. GraphXNET expanded by more than 20%, with AUC scores of 0.53 and 0.78 for 2% and 20% annotated datasets, respectively. Other methods differed only slightly: The AUC score of the bi-modality synthesis model increased by 4% from 1% to 10% labelled data, while the AUC score of the SSAC model increased by 3% to 100% labelled data.

Additional research can yield greater robustness to changes in the fraction of labelled data, as may effective training with fewer labelled data. As previously stated, the domains of bi and multi-modality image synthesis provide potential for further research. Furthermore, given the promise of the ACPL technique, it may be desirable to examine new methods for assessing confidence and uncertainty in order to achieve improved outcomes utilizing the traditional pseudo-labeling method. Another type of learning that may be worth investigating is active learning.

Conclusion

In this strategy, a medical professional is asked to name a few valuable unlabeled samples rather than using an existing dataset. Because the instances to be marked are chosen expressly based on the data they provide, this technique results in a less amount of named material required and, as a result, less time and effort in naming from specialists. Last but not least, more extensive studies are required to assess the usefulness of these models in a real-world clinical context with bigger patient groups.