Automatic diagnosis for vitiligo and other skin images in data-insufficient scenarios

• Main Author: Li,Yanling
• Other Authors: Kong Wai-Kin, Adams
• Format: Thesis-Doctor of Philosophy
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Computer and Information Science; Medicine, Health and Life Sciences
• Online Access: https://hdl.handle.net/10356/175955
• Institution: Nanyang Technological University

Automatic diagnosis methods can be applied to many fields in medical applications, such as skin disease severity diagnosis and organ cancer diagnosis. However, due to the tremendous efforts required to acquire medical images and concerns regarding patient privacy, large annotated medical datasets are rarely established. To obtain decent performance in medical applications, methods should be designed specifically for medical applications with insufficient data. This thesis presents three solutions to tackle the problem of insufficient data in the field of medical image processing. Currently, disease severity scores require dermatologists to estimate the percentage area of involvement, which is subjected to inter and intra-assessor variability. Automatic diagnosis methods can provide objective scores. However, previous studies have focused on either patch-wise skin or a particular body part while the full-body vitiligo segmentation, which can be converted to vitiligo clinical scores, was completely neglected. The proposed algorithms aim to address this gap by providing solutions for segmenting standard face images and body-part images. The results of these algorithms can be converted into a full-body vitiligo involvement percentage for clinical use. The first work proposes to segment standard vitiligo face images. To address the data scarcity issue, images from two different sources, the Internet and the proposed vitiligo face synthesis algorithm, are employed in training. Both synthetic and Internet images are used to train a CNN, which is modified from the fully convolutional network (FCN) to segment face vitiligo lesions. The results demonstrate that: 1) the synthetic images effectively improve segmentation performance; 2) the proposed algorithm achieves a 1.06% error for face vitiligo percentage estimation and 3) it outperforms two dermatologists and all the previous automated vitiligo segmentation methods, which were designed for segmenting vitiligo on pure skin. The second work aims to address standard body-part images in vitiligo. The first full-body vitiligo dataset, consisting of 1740 images, following the international vitiligo photo standard, was established in this work. Compared with patch-wise images, full-body images have more complicated ambient light conditions and larger variations in lesion size and distribution. Additionally, in some hand and foot images, the skin can be fully covered by either vitiligo or healthy skin. Previous patch-wise segmentation studies, which assume that the contrast between vitiligo and healthy skin is available in each image for segmentation, completely ignore these cases. This work exploits a tailor-made contrast enhancement scheme and long-range comparison, along with a novel confidence score refinement to improve the performance. Compared to the state-of-the-art method, the proposed algorithm reduces the average per-image vitiligo involvement percentage error from 3.69% to 1.81%, and the top 10% per-image errors from 23.17% to 8.29%. Our algorithm achieves 1.17% and 3.11% for the mean and max error for the per-patient vitiligo involvement percentage, respectively, which is better than an experienced dermatologist’s naked-eye evaluation. The third work focuses on few-shot object detection in a skin dataset, a CT dataset, and two common object datasets. Creating densely annotated data for model training requires significant effort, and data-insufficient categories are prevalent in medical images. Therefore, the ability to recognize unseen or rarely seen categories is crucial in these scenarios. A fully decoupled few-shot object detection algorithm is proposed to address the conflict between training classification and localization for unseen or rarely seen categories. Additionally, a training-free CLIP-based refinement module is introduced to improve detection performance. The proposed algorithm has been validated on a skin disease dataset, a public multi-organ CT dataset, and two common objects datasets. The results achieve state-of-the-art performance and underscore the importance of the fully decoupled approach in few-shot detection.