There are many types of weak signals in different organics of human bodies, and many devices can collect images of them, which can reflect different kinds of diseases, reactions, and chemical changes. With the development in medical, biology, and electronic areas, such different information plays an essential role in disease detection, health care, and other relative applications, especially the combination of different examination results. However, the large amount of signals or data collected from the human body needs more steps to clean the redundant information, extract the relative features, and classify the data for applications. So, the data can be simply divided into 1-dimension, 2-dimension, and 3 or more dimensions data before processing. To realize the application with high efficiency, computer-aided diagnosis (CAD) can serve as the assistant that has become a part of the routine clinical work for the detection of certain signals or images. With the help of CAD, radiologists or doctors can refer to the analysis results and then make final decisions. Currently, artificial intelligence (AI) methods also help the system learn to make diagnoses based on the input signal, and in some areas, these methods have shown advantages in classifying the abnormal tissues, tumors, bacteria, and different items under the images, distinguish the illness, monitor the status of patients and so on. It may need a large amount of data and frequent training to improve the accuracy. There also may be some noise or loss after the transportation of the signal from the in-body to the machine. Therefore, an efficient AI algorithm with fast training speed and hardware-involved methods are needed to realize these requirements.

This thesis presents a comprehensive system mainly based on extreme learning machine (ELM) for 1-D and 2-D bio-medical signal or image processing and diagnosis of certain symptoms and illnesses, because of the fast training and flexible structure of the framework built by ELM. For certain detection and diagnosis, a hardware and software co-design is presented to fulfill the requirements in noise and power consumption.

In exploring the system, some related works on different areas are designed with advantages compared to both conventional methods and state-of-the-art technology. Firstly, a hierarchical extreme learning machine (H-ELM) is employed in the EEG signal classification of motor imagery. This work is designed on a software level and processes all channels of 6-channel EEG signals, while other works may select 2-channel signals manually which are related to motor imagery. But on these conditions, this work still shows the advantages in accuracy and training time compared to conventional machine learning algorithms.

The second work proposes a novel hardware and software co-design based on Xilinx System Generator on the same EEG signal. This co-design separates the whole process into 2 levels and realizes high efficiency in different steps. For the special feature of ELM, the framework is randomly generated on software and trained by a dataset, and then it will be implemented on hardware for in-body detection and processing to reduce the noise and realize real-time detection.

The third work extends the objects from 1-D bio-signal to 2-D bio-image, and it shows an efficient system for breast cancer primary diagnosis based on H-ELM. The framework does not only show the advantages in accuracy, but it only requires general and small-size images with better performance, compared to conventional deep learning methods, which may require a much larger amount of dataset in training and training time. This work also transfers the algorithm into a system for primary diagnosis of breast cancer to distinguish whether the tumors are benign, malignant, or normal.

Finally, along with the works, a novel comprehensive ELM-based co-design system is built with different wearable sensors for different signals or images, software, and hardware processing units, and an optimized algorithm for certain formats and sizes of the input. The whole system is also realized with certain wearable devices and hardware with optimized algorithms to realize the real-time analysis of certain symptoms or illnesses.

The details of each work will be presented in the following relative chapter and all of these researches are compared with state-of-the-art technologies to determine the viability and efficacy of the newly developed systems. Besides, the potential future research scopes are also explored at the end of the thesis.