Feature-Based vs. Deep-Learning Fusion Methods for the In Vivo Detection of Radiation Dermatitis Using Optical Coherence Tomography, a Feasibility Study

Abstract

Acute radiation dermatitis (ARD) is a common and distressing issue for cancer patients undergoing radiation therapy, leading to significant morbidity. Despite available treatments, ARD remains a distressing issue, necessitating further research to improve prevention and management strategies. Moreover, the lack of biomarkers for early quantitative assessment of ARD impedes progress in this area. This study aims to investigate the detection of ARD using intensity-based and novel features of Optical Coherence Tomography (OCT) images, combined with machine learning. Imaging sessions were conducted twice weekly on twenty-two patients at six neck locations throughout their radiation treatment, with ARD severity graded by an expert oncologist. We compared a traditional feature-based machine learning technique with a deep learning late-fusion approach to classify normal skin vs. ARD using a dataset of 1487 images. The dataset analysis demonstrates that the deep learning approach outperformed traditional machine learning, achieving an accuracy of 88%. These findings offer a promising foundation for future research aimed at developing a quantitative assessment tool to enhance the management of ARD.

Keywords: Acute radiation dermatitis; Classification; Deep learning; Feature extraction; Machine learning; Optical Coherence Tomography.

Fig. 1

Flow chart of the study showing the two different classification procedures used. After image acquisition and segmentation, feature based ML with feature extraction was applied. Alternatively, feature HSV and pseudocolor images were created, followed by Deep Learning with late fusion

Fig. 2

A Digital photo of a patient presenting with two different grades of ARD. The top black box encloses an area of Grade 1 ARD whereas the lower black box outlines an area of Grade 2b ARD. B Three of the six OCT images acquired from the Grade 1 ARD area. (C) Another three OCT images from the Grade 2b ARD area

Fig. 3

A In vivo OCT image of human skin in the neck region. The automated algorithm segmented a given depth of skin (in this example 0.650mm in air (or approximately 0.450 mm in tissue), outlined by the green lines). The vertical red lines mark the portion of the image used (in this case the entire image). B The flattened segmented portion which was used to calculate intensity, texture, fractal, and scatterer size features at each distinct-window neighborhood (purple squares). C The same image as in (B), was used to determine the optical group velocity (GVD) dispersion based on the speckle resolution degradation at progressively increasing depths (yellow rectangles). D For each feature, a new image was created where the value of the features (in this example, scatterer size) was overlaid on the intensity as HSV color

Fig. 4

From each OCT imaging location, several feature images are created. First order intensity statistics (IS), GLCM second order statistics, group velocity dispersion (GVD) and fractal dimension (FD) images are created from the intensity image. The spectrally dependent properties of scatterer size (SS) and the bandwidth of the correlation of the derivative (COD BW) are extracted from the spectral content of each raw interferogram

Fig. 5

Training flow of the proposed multi-featured deep learning method. Each feature dataset passes through a separate ResNet101 neural network, and the resulting classes are combined into a feature vector which is then passed as the input to a traditional machine learning classifier to produce the final result