Google proposed a hybrid approach for medical imaging.

2019-12-10

Medical imaging is one of the most popular applications of artificial intelligence and machine learning. Computer vision algorithms are highly adept at detecting anomalies that experts may sometimes overlook, thanks to their precision. This is perhaps why, despite the relatively low adoption rate of artificial intelligence by global healthcare organizations (22%), the majority of practitioners (77%) still consider this technology to be crucial for the entire field of medical imaging. Data scientists have invested a significant amount of time and effort in developing artificial intelligence imaging models for healthcare systems. Recently, Google scientists detailed some of these models in a paper that was accepted at the NeurIPS conference held in Vancouver this week. In the book "Transfusion: Understanding Transfer Learning in Medical Imaging," Google scientists investigated the role of transfer learning in developing image classification algorithms. Transfer learning involves training machine learning algorithms in two stages. The first stage is retraining, where the algorithm is typically trained on benchmark datasets representing different categories. The next stage is fine-tuning, where it is further trained for the specific target task of interest. Finally, a pre-training step helps the model learn general features that can be reused for the target task, thereby improving its accuracy. Transfer learning is not the ultimate goal of artificial intelligence training techniques. In performance evaluation and a range of model architecture training tasks, including diagnosing diabetic retinopathy from retinal images and chest X-ray examinations for five different diseases, transfer learning did not have a "significant" impact on medical imaging tasks. During the tests, the team studied the extent to which transfer learning affects the features and representations learned by artificial intelligence models. They analyzed and compared hidden representations. Similarity scores were obtained between representations learned by models trained from scratch and those pre-trained on ImageNet for solving medical imaging tasks. Ultimately, the team concluded that for large models, representations learned from scratch tend to be more similar to each other than those learned from transfer learning, while for small models, there is a greater overlap in similarity scores between representations. To address these issues and others, the team proposed a hybrid transfer learning approach, where the complete model architecture is not reused. Instead, only a portion is recovered, while the rest is redesigned from scratch to better adapt to the target task. They stated that this approach retains most of the benefits of transfer learning while making the model design more flexible. "Transfer learning is a core technology in many fields," wrote Google research scientists Maithra Raghu and Chiyuan Zhang in a blog post. "Many interesting open questions still remain, and we look forward to addressing them in future work."

RECENT AI NEWS

RECENT AI TOOLS