Artificial Intelligence and Deep learning has sparked a lot of attention in the development of new medical image processing algorithms. Recently deep learning-based models have proven to be incredibly successful in a range of medical imaging responsibilities to help with detection and diagnosis of diseases. In spite of their success, deep learning models in medical image analysis are severely hampered by the scarcity of bulky, well-annotated datasets.

Both the accuracy of detection and diagnosis of cancers as well as other diseases in current clinical practise are dependent on the capability of individual clinicians, resulting in significant inter-reader variations in analysing and understanding medical images. To confront and resolve this clinical challenge, numerous computer-aided detectors and prognosis algorithms have been designed and evaluated, with the goal of assisting researchers and doctors with more efficiently reading medical images, also by making them quite accurate and thorough diagnostic decisions. The scientific reason for this approach would be to use computer-aided quantifiable image feature analysis as to help alleviate numerous harmful consequences in clinical practise, such as great variability in clinician expert knowledge, possibility fatigue of human experts, and a lack of adequate resources. Deep learning-related methodologies became the standard technology in the Computer aided field, with applications ranging from disease diagnosis to medical object recognition to image registration.

Among the various deep learning techniques, supervised learning was the first to be used in medical image analysis. Even though it’s been used effectively in several applications, the finite size within most medical datasets makes additional implementation of supervised models in several instances difficult. In comparison with conventional datasets in machine learning, a medical image dataset typically contains fewer images (less than 10,000), and only a low percentage of images are analysed by specialists in several cases. Unmonitored and semi-supervised learning strategies, that can create more labeled images for model optimization, discover constructive underlying patterns from unmarked image data, and produce pseudo labels for the unlabeled data, had also gained recognition in the last three years to overcome these limitations.

Deep learning could be classified as supervised, unsupervised, or semi-supervised based on whether or not labels from the training dataset have become prevalent. All image data are labeled in supervised methods, and the model is optimisedthrough using image-label pairs. The optimised model would then start generating a posterior probability score for every testing image in order to estimate its data point. Without markings, the model would then analyse and discover the correlations or concealed data structures in unsupervised learning. If just a subset of the training data is labeled, the algorithm achieves the input-output relationship from the labeled data and is bolstered by learning semantic and fine-grained functionalities from the unlabeled data. The said method of learning is known as semi-supervised learning.Massively evaluated recent advancements in unsupervised and semi-supervised learning could perform specific medical image functions with restricted annotated data. Widely known structures for both these two types of learning methodologies would be presented.

Finally, to summarise, there are three basic strategies for improving efficiency in medical image analysis which can be merged with various learning conceptual frameworks, which include algorithms, domain knowledge, as well as uncertainty estimation.