Zhongliang Lang1、2, Fan Zhang3, Bingxuan Wu3, Pengfei Shao3, Shuwei Shen4, Peng Yao5, Peng Liu4、**, and Xiaorong Xu1、3、4、*
Author Affiliations
1School of Biomedical Engineering, University of Science and Technology of China, Hefei 230026, Anhui, China2Department of Plastic Surgery, The First Affiliated Hospital of USTC (Anhui Provincial Hospital), University of Science and Technology of China, Hefei 230001, Anhui, China3Department of Precision Machinery and Precision Instrumentation, University of Science and Technology of China, Hefei 230027, Anhui, China4Suzhou Institute for Advanced Research, University of Science and Technology of China, Suzhou 215123, Jiangsu, China5Department of Electronic Science and Technology, University of Science and Technology of China, Hefei 230026, Anhui, Chinashow less
Fig. 1. Schematics of typical application scenarios of intelligence teleconsultation system for skin tumors built by our team. (a) Skin tumor self-screening; (b) tele-consultation
Fig. 2. Intelligence teleconsultation system for skin tumors. (a) System hardware composition; (b) schematic of AI-CPI module; (c) schematic of principle
Fig. 3. Software design of intelligence teleconsultation system for skin tumors. (a) User interface; (b) workflow under network condition
Fig. 4. Quantification experiment of projection accuracy of AI-CPI module. (a) Schematic of experiment; (b) projection errors at different working distances; (c) projection errors at different working angles
Fig. 5. Quantification experiment of resolution and color accuracy of AI-CPI module. (a) USAF 1951 resolution target plate captured by camera; (b) re-projected target plate; (c) changes in intensity of two regions of interest (ROI) along dotted line; (d) X-rite color card images taken by camera; (e) color card images after color calibration; (f) CIEDE2000 histogram before and after color correction
Fig. 6. Teleconsultation experiment of skin tumor using proposed system. (a) Subject and system at local site; (b) experienced expert is viewing images of skin tumor; (c) deep learning algorithm deployed in system diagnoses lesion as basal cell carcinoma through dermoscopic images; (d) expert confirms that diagnosis result of algorithm is correct, and draws annotations to discuss surgical plan; annotations projected on subject skin surface by proposed system are used to indicate (e) area of tissue removed and (f) area of skin flap reconstruction
Metric | RegNetY-800M | Dermatologist | p value |
---|
Balanced accuracy | 0.742±0.032 | 0.749±0.016 | p>0.05 | AUC | 0.899±0.016 | 0.926±0.001 | p>0.05 | Accuracy | 0.727±0.038 | 0.740±0.001 | p>0.05 | Specificity | 0.928±0.009 | 0.932±0.001 | p>0.05 | Precision | 0.730±0.049 | 0.752±0.011 | p>0.05 |
|
Table 1. Comparison of diagnosis results of algorithm in system with those of three dermatologists
Metric | Dermatologist without AI | Dermatologist with AI | p value |
---|
Balanced accuracy | 0.776±0.011 | 0.816±0.016 | p<0.05 | Accuracy | 0.740±0.001 | 0.780±0.016 | p<0.05 | AUC | 0.942±0.003 | 0.949±0.004 | p>0.05 | Specificity | 0.931±0.001 | 0.940±0.005 | p>0.05 | Precision | 0.751±0.012 | 0.807±0.015 | p<0.05 | Diagnostic time | (12.9±0.4)s | (10.0±0.8)s | p<0.05 |
|
Table 2. Comparison of diagnostic results of dermatologists in two groups with or without AI prompts