Skin cancer is among the most common cancers worldwide. Skin cancer screening relies primarily on visual inspection by dermatologists, and largely depends on their experience. However, imbalances in regional development have led to an uneven distribution of medical resources worldwide. Telemedicine is an effective approach to alleviate this dilemma, and a major branch of this field is teledermatology. Dermatologists can remotely view clinical images and medical histories of skin cancer patients in various ways and provide remote diagnosis and treatment suggestions. However, teledermatology relies on medical experts and network conditions, and patients who cannot access the Internet in remote areas cannot enjoy the convenience of remote consultation. Some portable devices based on automatic diagnostic algorithms have made up for some shortcomings of traditional teledermatology; however, because they can only judge the type of disease, these devices have limited clinical applicability. To compensate for the shortcomings of the current research, our team design and build a skin tumor artificial intelligence-enhanced teleconsultation system, which has both offline automatic skin tumor screening and online remote consultation and preoperative planning functions.

In this study, we build a skin tumor intelligent teleconsultation system with two typical application scenarios: 1) skin tumor self-screening without network conditions. Patients or doctors who lack experience at the local site can use the dermatoscope in the system to obtain images of skin lesions and then use the deep learning algorithm in the system to generate automatic diagnosis results. 2) Remote skin tumor consultation under network conditions. Remote consultation for skin tumors involves real-time interactions among inexperienced doctors at a local site and experienced medical experts at a remote site. First, the doctor at the local site uses a dermatoscope or ordinary camera in the system to obtain skin images of patients with skin tumors. Then, the deep learning algorithm in the system generates automatic diagnosis results based on the images. Finally, the remote expert confirms the disease diagnosis result and recommends a corresponding treatment plan based on network-transmitted images and algorithmic cues. For patients requiring surgery, both parties can use the system for preoperative planning. Instructive annotations drawn by remote doctors on the screen are transmitted over the network to the local site and projected onto the body surface of the patient in situ. To achieve the above functions, we first train a RegNetY-800M model based on 7-point dataset and deploy it on Raspberry Pi, also using a neural network computing accelerator to accelerate neural network computation, then use camera, laser projector, and beam splitter to form a co-axial projective imaging design that can project instructive annotations made by remote experts with high accuracy onto patient's body surface, finally design a visual software interface for easy use by doctors. To characterize the system, we first design a benchtop experiment to quantify the achievable accuracy of the system, then design a control experiment to verify whether the system can improve the applicability and efficiency of the traditional teleconsultation system. Finally, a clinical experiment is designed to verify the clinical applicability of the system.

The benchtop experimental results show that the maximum projection error of the system is less than 1.5 mm (Fig.4), and the CIEDE2000 value after color correction is less than 2, which can accurately restore colors in the scene (Fig.5). The control experimental results show that there is no significant difference between the algorithm in the system and dermatologists; dermatologists perform better with AI prompts. Clinical experimental results show that using this system for automatic diagnosis and remote consultation of skin tumors is feasible.

The intelligent teleconsultation system for skin tumors designed and built by our team has both automatic disease screening without a network and remote consultation with the network. The control experimental results show that the diagnostic results of the algorithm deployed in the system are comparable to those of dermatologists and that the algorithm can help dermatologists make more efficient and accurate decisions. The experimental results show that the system has clinical applicability. Compared to traditional remote consultation systems, this system has the following three advantages. 1) It has a wider application range. In remote areas without a network, the system can serve as a supplement to dermatologists, compensating for defects in which traditional teleconsultation systems cannot be used without a network. 2) Better performance. Automatic diagnosis results in the system can assist specialists in a more efficient and accurate screening of skin tumors. 3) Intuitive display. This system can project annotations made by remote experts onto body surfaces of patients with high precision, thereby making remote guidance in preoperative consultations and surgical planning processes more intuitive and precise. This system can help patients in areas with scarce medical resources perform early screening for various diseases, such as skin tumors.