The recent advancement in imaging technology, together with the hierarchical feature representation capability of deep learning models, has led to the popularization of deep learning models. Thus, research tends towards the use of deep neural networks as against the hand-crafted machine learning algorithms for solving computational problems involving medical images analysis. This limitation has led to the use of features extracted from non-medical data for training models for medical image analysis, considered optimal for practical implementation in clinical setting because medical images contain semantic contents that are different from that of natural images. Therefore, there is need for an alternative to cross-domain feature-learning. Hence, this chapter discusses the possible ways of harnessing domain-specific features which have semantic contents for development of deep learning models.
TopIntroduction
Semantic Medical Image Analysis (SMIA) might be the “next big thing” in scientific computing (SC) within the context of the healthcare industry basically because its analytics is contingent on semantics in its entirety (i.e. Intention, Meaning, and Context (IMC)). In SMIA resources such as processes, data, tools, document, device, people will be attended to. In the context of Semantic Computing (SC), SMIA is scoped around analytics, integration, description languages for semantics, and interfaces, etc. Additionally, applications that include biomedical systems, SDN, IoT, wearable computing, cloud computing, context awareness, mobile computing, big data, search engines, question answering, multimedia, and services will draw on SC in the 21st century to influence SMIA. The presentation in Table 1 shows what these applications will contribute to (or draw on) SC and aggregate impact on SMIA, which Machine Learning (ML) will benefit from. The sign (Í) show that IMC cannot be applied nor relevant, while (ü) shows it can be applied or relevant.
Table 1. Application and their contribution to SC and aggregate impact on SMIA
SMIA rely on Machine Learning Algorithms (MLA) to build implementable models using sample data. These sample data are “training data,” which provide the knowledge to train a model or algorithm to have its own information (i.e. experience) to predict outcomes accurately. This happens after training a model without necessarily programming it explicitly to perform the predictive task (Zhang, 2020). Where SC comes in is in the area of understandable insight and applicable intelligence. As such, SC complements MLA to handle the provision of intelligence and insight from data, which the traditional MLachine learning techniques cannot handle.
As is, there has been an explosion in the production of Medical Images (Med-I). This upsurge resulted from the use of Advanced Biomedical Data Collection Devices (ABDCD) that operate digitally with increased throughput (Shung et al, 1992; Razzak et al, 2018). Though, fueled by the current digital technological revolution, it came with its challenge. For example, in the pre-explosive era of Med-I production the volume of data that exist were easily annotated because they were few with manageable human annotation error. However, in this explosive era it has increasingly become difficult to manage this human error because of the huge volume of Med-I that are still annotated manually. The use of the traditional machine learning technique (i.e. algorithm) has fared well in the tasks of medical image analysis despite the massive volume of data produced by ABDCD that it handles per time. However, despite this successful shift in the paradigm of using computational machine learning models to solve the problem of annotation considering the huge volume of available Med-I data, there is still a disconnect regarding the semantic interpretation of the outcome of such annotation. This lack of interpretive disconnection still make it necessary to augment diagnostic results from High Performing Computational Models (HPCM) with human opinion before decisions are made.