| dc.description.abstract | Accurate early diagnosis of malignant skin lesions is critical in providing adequate and timely treatment. Unfortunately, initial clinical evaluation of similarly looking benign and malignant skin lesions can result in missed diagnosis of malignant lesions and unnecessary biopsy of benign ones. Image-guided tools capable of objectively discriminating malignant from benign skin lesions could potentially assist with the clinical evaluation of suspicious lesions and identify those patients benefiting the most from biopsy examination. The first part of this study tested the hypothesis that maFLIM-derived autofluorescence global features can be used in machine-learning models to discriminate malignant from benign pigmented skin lesions. Clinical widefield multispectral autofluorescence lifetime (maFLIM) dermoscopy imaging of suspicious skin lesions were acquired from 30 patients. Different pools of global image-level maFLIM features: multispectral intensity, time-domain bi-exponential, and frequency-domain phasor features, were extracted and compared. Ensemble combinations of quadratic discriminant analysis (QDA) models trained with phasor and bi-exponential features yielded a sensitivity of 84% and specificity of 90%. Simple classification models based on time-resolved autofluorescence global features extracted from maFLIM dermoscopy images have the potential to objectively discriminate malignant from benign pigmented skin lesions.
In the second part of this work, a deep learning model using Long Short-Term Memory (LSTM) networks is trained on the maFLIM images from the skin lesions. The model is trained at the pixel level and generates posterior probability maps for each image. An average area under the curve (AUC) of the ROC curves constructed on the median posterior probabilities of 0.82±0.04 is obtained. The third part of this study classifies oral cancer lesions using LSTM based deep learning models. Early diagnosis of oral cancer is critical as 70% of lesions are presented at their advanced stages, causing a drastic decrease in the five-year survival rate. The LSTM network is trained on the maFLIM images collected from the cancerous and healthy sites on the oral cavity. The pixel-level classification produced classification maps indicating the regions of malignancy for each patient. The overall sensitivity of 86.4% and specificity of 82% were obtained, with standard deviations less than 4%. Therefore, LSTM based deep learning models have the capability to classify benign and malignant skin and oral lesions at the pixel level. | |
