Deep Learning on 3D Cellular-resolution Tomogram

Overview

Skin biopsy is the main method to diagnose skin tumors, skin inflammation, and pigmented diseases. However, biopsy is an invasive method that can cause wounds and scars. Optical coherent tomography (OCT) technology is a fast, non-invasive, non-radioactive, and label-free imaging method. This technology generates real-time images of living tissue by detecting the variations in the refractive indexes of various components in soft tissues. Recently, there is a breakthrough progress that the newly designed ultrahigh resolution OCT can provide in vivo cellular resolution similar to histopathological sections in the high magnification. In our previous clinical trial "Early feasibility study: application of OCT imaging in dermatology" (approved by IRB of MacKay Memorial Hospital, no. 17CT062Be), it showed characteristic features of different skin inflammatory diseases and tumors can be distinguished successfully in tomograms. There were no adverse event or serious adverse event in this trial. Artificial intelligence technologies have been used widely in the image analysis in recent years. Hence, we aim to collect OCT tomograms of common skin inflammatory diseases, skin tumors, and pigmented diseases, and compare with normal skin for machine learning. We expect the integration of tomograms with deep learning artificial intelligence may assist identifying histological features in these images and provide new alternative way for non-invasive diagnosis in dermatology.

Introduction Optical coherent tomography (OCT) technology has been widely used in medical practice, such as ophthalmology. The application in dermatology is slowly progressed until the marked improvement of resolution recently. One of the newly designed OCT devices using in this study is based on the research and development of Professor Sheng-Lung Huang of National Taiwan University. The light source was made with original glass-covered crystalline fiber which has successfully provided sub-micron resolution on the skin, which is better than the traditional 5-10 micron resolution of high-definition OCT. This new OCT system (ApolloVue™ S100 image system, Viper1-S003, Apollo Medical Optics) has been used in this previous clinical trial "in vivo OCT images of different skin diseases" without adverse events. OCT images of different skin diseases collected in that trial were compared with HE-stained pathological sections. They provided useful information to physicians. The risk-benefit assessment of this clinical trial is the same as expected. The risk is low in clinical use, and both for the operators and the subjects. In recent years, the application of artificial intelligence technology in the analysis of tissue classification of medical images is rapidly developing. Therefore, we are going to use deep learning technology to improve the interpretation of OCT images to help the subsequent diagnosis of skin diseases. Inclusion criteria Experimental group: 1. Adults aged 20 years or older 2. Non-treat lesion of epidermal inflammatory disease: dermatitis and psoriasis: 300 participants. 3. Benign tumors: seborrheic keratosis and nevus: 300 participants 4. Malignant tumors: actinic keratosis (AK), melanoma, basal cell carcinoma (BCC), Bowen's disease, squamous cell carcinoma (SCC), and extramammary Paget's disease (EMPD): 100 participants 5. Pigmented diseases: solar lentigo, melasma, and vitiligo: 300 participants Control group: The healthy face (exposed site) and inner forearm (unexposed site) skin of epidermal tumors and pigmented diseases of the above experimental group were used as a control group, excluding epidermal inflammatory diseases, 700 participants in the control group were expected. Exclusion criteria Experimental group: 1. Minors aged under 20 years 2. Suspected a transcutaneous infectious disease, including infections such as bacteria, fungi, viruses, and parasites. 3. All skin tumors that are in the subcutaneous tissue 4. All skin lesions are open wounds 5. All skin lesions are in a location that is difficult to scan 6. Not willing to cooperate with methods and related procedures of this study 7. Vulnerable populations, including prisoners, pregnant women, handicapped, mentally disabled, known AIDS patients, and homelessness Control group: 1. Minors under 20 years of age. 2. Epidermal inflammatory disease 3. Suspected a transcutaneous infectious disease, including infections such as bacteria, fungi, viruses, and parasites. 4. Individuals who have a systemic skin disorder. 5. Individuals who have a history of severe skin condition 6. Individuals with surgeries/cosmetic surgeries/micro cosmetic surgery (eg. cosmetic injections and/or laser etc.) on healthy skin at face and inner forearm in last 3 months and a physician determine the surgery will affect outcome of the OCT images. 7. Not willing to cooperate with methods and related procedures of this study 8. Vulnerable populations, including prisoners, pregnant women, handicapped, mentally disabled, known AIDS patients, and homelessness Deep convolutional neural network (DCNN) was used to mark tissue and lesions in OCT images. When training DCNN models, transfer learning strategies will be used to fine-tune the parameters from pre-trained models that contain a lot of image knowledge, such as GoogLeNet, rather than training the models from scratch. This method retains the low-level image knowledge common to natural and medical images, and significantly reduces the time to train the model. During the training process, the parameters of the first few layers that store the low-order image knowledge in the model are fixed, and the parameters of the subsequent layers of the model are changed by the back-propagation algorithm. Finally, a layer of linear classifier is added to the end of the DCNN to determine the type / size of the symptoms in the input image.