Machine Learning & Deep learning

Artificial Intelligence (AI) is defined as "the science and engineering of making intelligent machines" or "the field of computer science that simulates intelligent behavior using computers." Recently, there have been cases of outstanding performance by applying AI to various fields such as medical, electronic communication, and distribution system.

Machine learning is defined as a 'research field that gives a computer the ability to perform actions not defined by codes' in a specific approach to the field of AI, and learn and design a model that fits the label values for each data. Examples of algorithms that are widely used include SVM, Naïve Bayes, Decision Tree, etc., which are applied to problems such as speech recognition and spam classification.

Data is the key to build a model, and it can be divided supervised learning and unsupervised learning due to the presence or absence of labeled data. The supervised learning uses the label data in training data, and a representative algorithm is CNN (Convolution Neural Networks), RNN (Recurrent Neural Networks), DNN (Deep Neural Network), SVM (Support Vector Machine). In contrast, unsupervised learning does not use labeled data and analyzes the rules and features of data using techniques such as clustering, spatial transformation techniques such as auto-encoder, and density estimation using kernels. Other learning methods include semi-supervised learning methods that use a mixture of labeled and non-labeled data, and reinforcement learning methods that reward the model. Therefore, the more "high quality" data, the better the performance model. Here, the meaning of "good quality" can be said to the extent that it matches the distribution of data generated in the real environment. In other words, unbiased sampling is important. The sampled data is distributed for training, validation, and experimentation and cross-validates the trained model to select the optimal model.

Recent advances in computing capabilities have enabled GPUs (Graphics Processing Unit) to be applied to machine learning. With the large number of CUDA cores performing large amounts of operations at the same time/at once, deep learning based on neural networks, which has not been noticed before but nowadays it has attracted much attention. Deep learning is a neural network-based algorithm based on the human brain, and is a field of machine learning. By applying the GPU to the artificial neural network, a large number of nodes can be processed quickly, enabling a deeper and larger network. Algorithms such as Resnet, PSP Net (Pyramid Scene Parsing Net) and VGG-16 based on CNN, which utilize context of input data and significantly reduce parameters than Fully Connected Networks, have led to rapid advances in image processing areas such as Semantic Segmentation and Object Detection. In addition, the field of natural language processing was developed through RNN, a model that predicts future state by using information of current state and past state, and AI Speaker, Chat Bot, etc. with improved performance have been developed.

MedySapiens's Application Cases

Medysapiens developed a deep learning model-based solution, it segments three major arteries which are LAD (Left Anterior Descending artery), LCX (Left Circumflex Artery) and RCA (Right Coronary Artery), and detect lesion from multiple views of coronary X-ray angiography. The solution is based on a deep learning model that classifies syntax scores, risk indication indexes for coronary artery disease, and grades, which will effectively assist cardiologists in assessing and diagnosing coronary artery disease (CAD).

Medysapiens is developing a model for predicting the pathogenicity with the disease of the gene mutation(variant) identified from the NGS sequencing data of the gene. In addition to direct information such as nucleotides / amino acid sequences, structural information to determine the protein function (function), other in-silico prediction, etc., mutation pathogenicity and associated with high input and is developing a deep learning-based model accordingly.