Research on Dermatological Diagnosis System Based on Convolutional Neural Network

Research on Dermatological Diagnosis System Based on Convolutional Neural Network

1. Project Introduction

1.1 Project Description

This project uses a convolutional neural network and other methods to identify skin diseases based on clinical images and proposes a complete Dermatological Auxiliary Diagnosis System. The database of this project is from Xiangya Hospital of Central South University. This database is the largest clinical image data collection of psoriasis in the world.

1.2 Project Goal and Significance

In recent years with the rapid development of computer technology, computer aided diagnosis (CAD) has also made a series of progress, especially in the field of medical imaging. Practice has proved that CAD has played a significant positive role in improving diagnostic accuracy, reducing missed diagnosis and improving work efficiency. Therefore, this project has great significance for improving the current medical level and alleviating the pressure of dermatologists.

2. Main Method

The algorithm model initially adopts the Inception v3 model. Based on the original model, we use the idea of migration learning to obtain the initial weights in ImageNet pre-training and then fine-tune the network to achieve the purpose of effectively distinguishing different kinds of skin diseases.

Activation Map of Psoriasis Images in Inception V3 Network

3. Database

Step one was the collection. All images were collected by dermatologists with informed consent of patients from Xiangya hospital. This prospective study was approved by the Ethics Committee of Xiangya Hospital, Central South University and informed consent was obtained from all participants. To achieve superb quality of images, a professional digital camera (Canon, Resolution: 350dpi) was used to capture each visual feature of skin lesion with maintaining sufficient illumination. In addition, images of different zoom levels were also collected for capturing both spatial distribution characteristics and fine features such as texture. Finally, XiangyaDerm-Pso9 containing 8021 images was accumulated by selecting nine disorders (900 images of psoriasis). Fig. 1 shows a set of sample images in XiangyaDerm-Pso9. For better generalization performance, the diversity of dataset is necessary. These images are clinical dermatological images containing multiple body parts of patients.

Database Sample Map

Step two was data filter and annotation. After the data collection was completed, we filtered out the unqualified images and labelled the remaining ones. We removed four categories of images to obtain a clean dataset during data preprocessing: Case 1: skin lesions were obscured or altered by visible topical treatments or any other colored residues, which can incur deep adverse impact on the training process. Case 2: special parts, such as fingers, where lesions are smaller and appear darker than normal. Case 3: skin lesions covered by hair or any other visual elements, which makes it hard to extract interpretable images. Case 4: Excessive exudate, which leads to loss surface appearance and disease-specific texture. Subsequently, three professional dermatologists who have been engaged in dermatology for more than 10 years from Xiangya Hospital annotated each image according to the corresponding medical record and pathology results to generate the final standardized experimental dataset. The data filter and annotation process is shown as following.

Schematic Diagram of Data Processing

4. APP Design

The system establishes two methods of image acquisition, one is to shoot directly through the mobile phone and the other can be called from the local library in the mobile phone. Pictures obtained through different channels will be uploaded to the local server. After receiving the pictures from the mobile phone, the local server will be imported into the previously designed classifier for classification. Before getting the result document, the phone will continuously request the result document from the server until the image analysis is completed. Finally, after receiving the result document, the phone prints it on the result display interface.

APP Partial Screenshot

5. Project Progress and Existing Achievements

①Competition Results

The project participated in the artificial intelligence creative competition of the China University Computer Contest with the work AIMADS (skin disease part) and won the first prize of the National Finals. AIMAS (Artificial Intelligence Medical Aided Aiagnosis System) can be used for the auxiliary diagnosis of common skin diseases (compatible with Android and IOS). The patient or doctor can upload the affected picture to the server on the client and the server can pass the present. The deep convolutional neural network model completed by the team design training analyzes the uploaded data and displays the diagnosis results in the APP. For the diagnosis of dermatological diseases, the classification of common skin diseases (Lixia Lichen, Lupus Erythematosus, Basal Cell Carcinoma, Squamous Cell Carcinoma, Normal Skin, Pemphigus, Psoriasis, Seborrheic Keratosis) can be directly given and given Out of confidence.

AIMADS Skin Diagnosis of Intelligent Diagnostic System

② Thesis Results

1.The Application of Deep Learning in the Risk Grading of Skin Tumors for Patients Using Clinical Images

According to diagnostic criteria, skin tumors can be divided into three categories: benign, low degree and high degree malignancy. For high degree malignant skin tumors, if not detected in time, they can do serious harm to patients’ health. However, in clinical practice, identifying malignant degree requires biopsy and pathological examination which is time costly. Furthermore, in many areas, due to the severe shortage of dermatologists, it’s inconvenient for patients to go to hospital for examination. Therefore, an easy to access screening method of malignant skin tumors is needed urgently. Firstly, we spend 5 years to build a dataset which includes 4,500 images of 10 kinds of skin tumors. All instances are verified pathologically thus trustworthy; Secondly, we label each instance to be either low-risk, high-risk or dangerous in which Junctional nevus, Intradermal nevus, Dermatofibroma, Lipoma and Seborrheic keratosis are low-risk, Basal cell carcinoma, Bowen’s disease and Actinic keratosis are high-risk, Squamous cell carcinoma and Malignant melanoma are dangerous; Thirdly, we apply the Xception architecture to build the risk degree classifier. The area under the curve (AUC) for three risk degrees reach 0.959, 0.919 and 0.947 respectively. To further evaluate the validity of the proposed risk degree classifier, we conduct a competition with 20 professional dermatologists. The results showed the proposed classifier outperforms dermatologists. Our system is helpful to patients in preliminary screening. It can identify the patients who are at risk and alert them to go to hospital for further examination.

Procedure of risk grading system for skin tumor based on deep learning

2.Smart Identification of Psoriasis by Images Using Convolutional Neural Networks: A Case Study in China (Not published)

Background: Psoriasis is a chronic inflammatory skin disease which holds a high incidence in China. However, professional dermatologists who can diagnose psoriasis early and correctly are insufficient in China, especially in the rural areas. A smart approach to identify psoriasis by pictures would be highly adaptable countrywide, and could play a useful role in early diagnosis and regular treatment of psoriasis. Objectives: Design and evaluation of a smart psoriasis identification system based on clinical images (without relying on a dermatoscope) that works effectively similar to a dermatologist. Methods: A set of deep learning models using convolutional neural networks (CNNs) were explored and compared in the system for automatic identification of psoriasis. The work was carried out on a standardized dermatological dataset with 8021 clinical images of 9 common disorders including psoriasis along with full electronic medical records of patients built over the last 9 years in China. A two-stage deep neural network was designed and developed to identify psoriasis. In the first stage, a multi-label classifier was trained to learn the visual patterns for each individual skin disease. In the second stage, the output of the first stage were utilized to distinguish psoriasis from other skin diseases. Results: The area under the curve (AUC) of the two-stage model reached 0.9810.015, which outperforms a single-stage model. And the classifier showed superior performance (missed diagnosis rate: 0.03, misdiagnosis rate: 0.04) than 25 Chinese dermatologists (missed diagnosis rate: 0.19, misdiagnosis rate: 0.10) in the diagnosis of psoriasis on 100 clinical images. Conclusions: Using clinical images to identify psoriasis is feasible and effective based on CNNs, which also builds a solid technical base for smart care of skin diseases especially psoriasis using mobile/tablet applications for teledermatology in China. Keywords: psoriasis, convolutional neural network, skin disease dataset, clinical image, classifier

Prospect of Psoriasis Intelligent Identification System

3.XiangyaDerm: A Clinical Image Dataset of Asian Race for Skin Disease Aided Diagnosis (Not published)

Skin disease is a quite common disease of human beings, which has been found in all races and ages. It seriously affects people's quality of life or even endangers people's lives. In this paper, we propose a large-scale, Asian dominated dataset of skin diseases with bounding box labels, namely Xiang-yaDerm. It contains 107,565 clinical images, covering 541 types of skin diseases. Each image in this dataset is labeled by professional doctors. As far as we know, this dataset is the largest clinical image dataset of Asian skin diseases used in Computer Aided Diagnosis (CAD) system worldwide. We compare the classification results of several advanced Convolutional Neural Networks (CNNs) on this dataset. InceptionResNetV2 is the best one for 80 skin disease classification whose Top-1 and Top-3 accuracies can reach 0.588 and 0.764, which proves the usefulness of the proposed benchmark dataset, and gives the baseline performance on it. The cross-test experiment with Derm101 shows us that the CNN model has a very different test effect on different ethnic datasets. Therefore, to build a skin disease CAD system with high performance and stability, we recommend to establish a specific dataset of skin diseases for different regions and races.

Experimental confusion matrix for six categories of skin diseases

Copyright Announcement

Please respect our original works. When reproducing, please mark the original source of the article and the author's information. If you have any questions or ideas to communicate with us, please contact us by email or telephone at the bottom of the page.