Machine Learning Assisted OCT For Clinical Diagnosis

Optical coherence tomography(OCT)is an imaging modality in biomedical research since 1990 s.The unique features of OCT make it the gold standard in diagnosis of fundus diseases in ophthalmology and the world shows a keen interest to apply it in other clinical areas.This thesis focused on the application of OCT in dentistry and dermatology.Several OCT systems were built and researches were conducted.In the process of analyzing clinical images,we found that speckle noise would degrade the image quality.Considering the output of the supercontinuum source pumped by noise-like pulses is random over time,we built a spectral domain OCT system based on the noise like source.With this system,we successfully removed the speckle noise in the image.The SNR,CNR,and ENL of denoised image were improved by 3.3 d B,2 d B,and 5.2,respectively.The XCOR between original image and denoised image was 0.794.Compared to the traditional methods of speckle reduction,the system required by our method is simpler and more stable.Port-wine stain(PWS)is the congenital capillary malformation of skin.The diagnosis of port-wine stain relies heavily on the subjective judgment and personal experience of the doctor.We use machine learning method to analyze the port-wine stain images.In our work,220 OCT images from clinical patients were analyzed.Accuracy of92.7%,sensitivity of 92.3%,and specificity of 93.8% were achieved with the full feature set.Accuracy of 92.7%,sensitivity of 94.9%,and specificity of 87.5% were achieved with the simplified feature set.These studies will contribute to a more objective and accurate clinical diagnosis of PWS patients.Another problem is that,most patients respond positively to the photodynamic therapy,but a few patients rarely respond to the photodynamic therapy.We try to use anomaly detection method to automatically identify the patients who rarely respond to the photodynamic therapy.Accuracy of 86%,sensitivity of 91%,and specificity of 84%were achieved.The method could be used to predict the response of patients and choose the appropriate treatment plan for patient in clinical.For clinical application,a swept source OCT system was built to be used in dentistry.We did some polit study in the process of filling the tooth.First,the cleaning effect of the rotted material was evaluated.Second,the evaluation of resin shrinkage and effectiveness of the doctor's operation was performed.During the study of PWS diagnosis,the system was a time-domain OCT system with relatively low signal to noise ratio,slow imaging speed and bulky size,severely restricting further research on clinical application.Therefore,we have designed and built a portable,high-speed,and handhold swept source OCT system with weight of 7.6kg,dimensions of 38 cm × 32 cm × 17 cm,SNR of111.7 d B,and speed of 100160 Alines/s.The new handhold system was used in image collection in clinic.Our studies show that OCT has high potential in the application of clinical dentistry and dermatology.Furthermore,combining OCT with machine learning is of great value in the diagnosis of clinical diseases,determination of treatment plans,and evaluation of treatment effects.