Computer Vision News - July 2022

42 MedTech Research Hello again! I hope that this month you enjoy the coverage of themost important conference in Computer Vision, if so, Ralph has you covered! Meanwhile, don’t forget medical imaging. In this month, I will present you a paper from Laith Alzubaidi et al, from Queensland University in Brisbane, Australia who kindly provided permission to use parts of the paper. As a reminder, it is always important to read the original paper and get your own appreciation of what is included there. I hope this month’s article will make this journey easier! Introduction Deep learning (DL) methods , as you probably already know, allow computational models composed of multiple processing layers to learn representations of data with multiple levels of abstraction. DL methods have been extensively used and achieved improved results compared to previous research in many fields, such as visual object recognition, speech recognition and object detection. It is known that the performance of DL depends on a huge amount of data for representation learning. For medical images, such as ophthalmology or pathology slides, the available data are often not sufficient to obtain a good model by using ImageNet dataset . Moreover, medical image analysis struggles from a lack of sufficient data for training DL models. Transfer learning helps to obtain accuracy for multimedia tasks. It was recently proven to help pre-trained models with medical grey-scale images, but as medical imaging tasks are becoming more prevalent, transfer learning is also becoming more difficult. This will improve the performance of machine learning algorithms when used on limited- data clinical problems, such as that of diseases diagnosis and prediction. In this paper you’ll discover a proposed newmodel for addressing the previous shortcomings of medical image classification. The model is called MedNet and is two-fold: a Gray-MedNet and a Color- MedNet . Both can be trained using a publicly available 3M medical image dataset of each version coming from multiple sources: MRI, CT, X-Ray, and PET, among many others. MedNet and perspective The lack of training data is a problem with modern deep CNNs. The common solution in state-of-the-art (SoTA) is to use transfer learning . Transfer learning relies on the knowledge of a specific task and uses it to address both the target task and other correlated tasks with Vision Transformers in Medical Computer Vision