Medical Image Analysis with Deep Learning under Limited Supervision Restricted; Files Only

Xiaoyuan Guo (Spring 2022)

Permanent URL:


Medical imaging plays a significant role in different clinical applications such as detection, monitoring, diagnosis, and treatment evaluation of various clinical conditions. Deep learning approach for medical image analysis emerges as a fast-growing research field and has been widely used to facilitate challenging image analysis tasks, for example, detecting the presence or absence of a particular abnormality, diagnosis of a particular tumor subtype. However, one important requisite is the large amount of annotated data for supervised training, which is often lacking in medicine due to expensive and time-consuming expert-driven data curation process. Data insufficiency in medical images is also limited by healthcare data privacy requirements, which leads to barriers in the usage of deep learning methods across institutions. This thesis focuses on facilitating the applications of deep learning approaches to solve automatic medical image analysis tasks efficiently under limited supervision. Three situations are in consideration: (1) no annotated data; (2) limited annotated data; (3) curation of additional annotated data with minimal human supervision. The research covers multiple medical image modalities starting from fluorescence microscopy images (FMI), histopathological microscopy images (HMI) to mammogram images (MG), computed tomography (CT), chest radiographs (X-ray). A variety of medical image related tasks have been researched, including clumped nuclei segmentation in FMI, clustered liver steatosis segmentation in HMI, segmentation and quantification of breast arterial calcifications (BAC) on MG, out-of-distribution (OOD) detection for medical images, shift data identification from unseen external datasets, image retrieval in external datasets with OOD-awareness and accurate multi-label medical image retrieval. Due to the reality of limited supervision in medicine, unsupervised, weakly-supervised, and supervised deep learning techniques have been investigated in this thesis to solve the medical tasks under different situations. The diversity and concreteness of the thesis can be a guide to facilitate the efficient usage of deep learning approaches in future medical image analysis with minimal cost.

Table of Contents

1. Introduction

1.1 Motivation

1.2 Research contributions

1.2.1 Medical image segmentation with limited supervision

1.2.2 Medical OOD identification with limited supervision

1.2.3 Medical dataset curation with limited supervision

1.2.4 Medical image retrieval with limited supervision

1.3 Organization of the thesis

2.Background & Related Work

2.1 Medical image segmentation with limited supervision

2.2 Medical OOD identification with limited supervision

2.3 Medical dataset curation with limited supervision

2.4 Medical image retrieval with limited supervision

3.Medical Image Segmentation with Limited Supervision

3.1 Clumped nuclei segmentation

3.1.1 Contribution

3.1.2 Method

3.1.3 Experiments

3.1.4 Conclusion

3.2 Liver steatosis segmentation

3.2.1 Contribution

3.2.2 Method

3.2.3 Experiments

3.2.4 Conclusion

3.3 Breast arterial calcifications (BAC) segmentation

3.3.1 Contribution

3.3.2 Method

3.3.3 Experiments

3.3.4 Conclusion

3.4 Discussions and future works

4.Medical OOD Identification with Limited Supervision

4.1 Medical novelty identification

4.1.1 Contribution

4.1.2 Method

4.1.3 Experiments

4.1.4 Conclusion

4.2 Generic medical anomaly detection

4.2.1 Contribution

4.2.2 Method

4.2.3 Experiments

4.2.4 Conclusion

4.3 Discussions and future works

5.Medical Dataset Curation with Limited Supervision

5.1 Contribution

5.2 Method

5.2.1 Problem statement

5.2.2 Formulation and notation

5.2.3 Shift identification

5.2.4 Anomaly detection

5.2.5 Shiftness quantification

5.2.6 Dataset quality measurement

5.3 Experiments

5.3.1 Datasets

5.3.2 Anomaly detectors in use

5.3.3 Experimental setup

5.3.4 Results

5.4 Conclusion

5.5 Discussions and future works

6.Medical Image Retrieval with Limited Supervision

6.1 Outlier-sensitive radiology retrieval

6.1.1 Contribution

6.1.2 Method

6.1.3 Experiments

6.1.4 Conclusion

6.2 Multi-label medical image retrieval

6.2.1 Contribution

6.2.2 Method

6.2.3 Experiments

6.2.4 Conclusion

6.3 Discussions and future works


8 Conclusion and Future Works


About this Dissertation

Rights statement
  • Permission granted by the author to include this thesis or dissertation in this repository. All rights reserved by the author. Please contact the author for information regarding the reproduction and use of this thesis or dissertation.
  • English
Research Field
Committee Chair / Thesis Advisor
Committee Members
Last modified Preview image embargoed

Primary PDF

Supplemental Files