Towards Robust Representation Learning for Computer-Aided Diagnosis under Weakly Supervised Information
The rapid development of artificial intelligence (AI) enables computers to identify a variety of diseases and abnormal anatomies from medical images, which can greatly reduce the radiologists' screening workload and realize early diagnosis and treatment. However, a large number of training data with high-quality annotation is required in today's computer-aided diagnosis (CAD) systems. It is very costly and time consuming, since annotators are normally specialized radiologists with related medical knowledge and manual annotation is prone to error when performed under a time restriction or by an entry-level radiologist. Unlabeled data and data with noisy labels, so called, label scarcity and label noise problems, are commonly encountered in medical image analysis. To tackle these two intractable problems, this proposed project will use machine learning (ML) technologies and develop robust, efficient and automated diagnosis algorithms which can be applied to identify diverse diseases (e.g. brain tumor, skin cancer, and pulmonary nodule). We will verify our proposed methods on a series of public datasets, such as MICCAI BraTS, MICCAI iSeg2019, ChestX-ray14 and ISBI CHAOS. We believe that this project will reduce the demands of annotated medical data, decrease the costs of manual screening, and prompt the development of smart healthcare.
Based on our solid research experience in both ML and healthcare, we investigate to use low-quality data and develop novel ML algorithms to achieve robust, efficient and automatic medical diagnosis. To address the problem that new data are unlabeled, or the labeled data may be too few, we investigate to utilize previous data with annotations, and design transfer learning algorithms which intelligently apply knowledge learned previously to help identify a new disease. We also theoretically analyze and model the relation between noisy and clean labels to handle the issue of label noise, which further improves the robustness of designed models. In addition, we apply our proposed algorithms to realize various medical applications, such as detection, diagnosis, and segmentation. We hope that our designed model can provide reasonable medical interpretation for doctors, helping them better understand the functioning mechanism of intelligent medical diagnosis.
- The candidates should have a solid background in mathematics, such as optimization, statistics and functional analysis.
- The candidates should be proficient with programming languages (Python) and deep learning tools (Pytorch).
- The candidates should have strong academic background in machine learning and medical image analysis, such as publications in NeurIPS/MICCAI/TPAMI/TIP.