One-class machine learning classification of skin tissue based on manually scanned optical coherence tomography imaging

Selection of features for classification.

We used U-Net as the feature extractor as described in the SVM for One-Class Classification section. Here, we evaluated different feature selection strategies. To establish the feature vector, we forward propagated the input image through U-Net to the layer before the segmentation layer (Fig. 3). The selected layer had N neurons (N = 16) and N activation values ​​were provided. For each OCT image patch with 32 Ascans, we averaged the activations of all pixels without distinguishing between pixel types to establish a feature vector. X_allIn addition, the obtained X_e By averaging the activation values ​​of the obtained epidermal pixels determined by U-Net X_d Average the activation values ​​of the dermal pixels determined by U-Net. moreover, X_edit Concatenate X_e When X_d: X_edit=[X_e;X_d]When computing the feature vector, we excluded pixels classified as “air” by U-Net because they were overwhelmed by noise. Stratum corneum pixels were not considered because few pixels fall into the stratum corneum. Since pixels belonging to different categories (epidermis and dermis) have different characteristics, we extracted feature vectors for different groups of pixels. To evaluate the feature selection strategy, we used 50% of the images in the data set (1116 image patches) to create a feature vector. X_all, X_e, X_d When X_editfor SVM training. For each set of training vectors, a one-class SVM classifier (SVM_allSVMs_eSVMs_dand SVMs_edit) with a Gaussian kernel function and a specific outlier ratio.

We validated the classification accuracy using the remaining image patches from the dataset. We extracted feature vectors from these images, fed the feature vectors to a trained classifier, determined the tissue type according to the classifier output, and determined the classification accuracy. Figure 5a shows the one-class classification accuracy obtained comparing the SVM classification to the ground truth (100% normal example). Figure 5a shows that a classifier trained with a higher outlier ratio tends to classify a larger percentage of test examples as anomalous.

(a) Classification accuracy validated using data obtained from normal skin. If the classifier was trained with different outlier proportions. (b) ROC curves for different classifiers when validated using a dataset consisting of images of normal skin and computer-synthesized abnormal images.

Additionally, we created a dataset using image patches (1116 patches) that were not used for SVM training. The dataset had 4464 image patches including OCT data of normal skin, synthetic BCC data, synthetic SCC image data and synthetic data with DEJ disruption.Aberrant images representing BCC were simulated by starting at random depths within the dermis and reducing the magnitude of the OCT signal to 75% of the original value.^{twenty one}We simulated anomalous images of SCC characterized by discrete bright areas under the skin surface by enhancing the signal magnitude by 25% within randomly selected regions under the skin surface.^{twenty two}We also created images in which the dermal-epidermal junction (DEJ) was disrupted by normalizing the individual Ascans in the image using the average depth profile. For each image patch, we created a feature vector (X_all, X_e, X_dWhen X_edit ). We labeled them as ‘normal’ when obtained from normal OCT data and ‘abnormal’ when obtained from synthesized abnormal data (BCC, SCC and DEJ disruption). We fed these feature vectors to a classifier (trained with an outlier rate of 8%). We obtained the operating characteristic (ROC) curves shown in Fig. 5b and listed the area under the curve values ​​for various classifiers in Table 1. We also compared the predictions provided by the classifiers with the ground truth and summarized the prediction accuracies in Table 1. Results The feature vector (X_edit=[X_e;X_d]) outperformed other feature vectors.Therefore I chose to use X_editfor subsequent classification of experimental data. Using a MacBook Pro computer (Apple M1 CPU and 8 GB RAM) and Matlab R2022a, it takes about 0.1 s to extract feature vectors from an image patch with 32 Ascans by following the steps shown in Fig. 4b. It takes about 0.01 seconds for the SVM classifier to make a prediction.

Spatial decomposition tissue classification based on one-class tissue classification

To demonstrate how a one-class classifier enables spatially resolved tissue classification, we scanned a fiber optic OCT probe from the thumb skin to the nail plate of a healthy subject. The resulting image is shown in Fig. 6a. The left side of the image corresponds to the skin and the right side of the image corresponds to the nail plate. OCT signals obtained from nails were considered abnormal, unlike skin.For a given abscissa image patch, we extracted features from epidermal and dermal pixels and established these features by concatenating X_edit. Pretrained One-Class SVM Classifier Using SVM_edit, we were able to obtain prediction scores at different spatial positions (Fig. 6b, black curve). To determine the boundary between normal skin and abnormal (nail plate), we filtered the SVM prediction scores (wavelet domain threshold) and obtained the first-order difference of the filtered SVM prediction scores (Fig. 6b red curve). The peak position of the red curve corresponds to the boundary between normal and abnormal tissue where the SVM prediction score changes abruptly. Boundary locations are shaded in red color in Fig. 6c, where a one-class SVM using features extracted from both epidermis and dermis enables spatially resolved tissue classification and tissue boundary detection. suggests that it has happened.

(a) OCT images obtained by scanning the junction between the skin and the nail plate of a healthy subject with a fiber optic probe. (b) spatially resolved SVM prediction scores (black curve), and first-order differences of filtered prediction scores (red curve). (c) Boundary between skin and abnormal tissue (red shading) identified by one-class SVM classification.

A pilot patient study.

In a pilot clinical imaging experiment, we imaged a 72-year-old male patient with a biopsy-confirmed BCC (nodular type) located in the left jaw (Fig. 7a). To benchmark the tumor OCT images against normal skin images, we acquired normal skin OCT images from two different locations on the patient’s forearm (Fig. 7b,c). In his OCT images obtained from the patient’s normal skin, the first layer of skin (stratum corneum) is thin and bright, followed by a decrease in brightness of the epidermis, and the DEJ is clearly visible. Beneath that is the dermis, where the signal decreases with depth. The tumor was scanned according to trajectories 1–4 shown in Fig. 7d and the resulting images are shown in Fig. 7e–h. Compared with normal skin of the same patient, images obtained from the tumor show disruption of the DEJ and reduction of his OCT signal amplitude starting from the upper dermis. We also scanned the area immediately adjacent to the circle drawn by the surgeon, following trajectories 5–8 shown in Fig. 7d. The resulting images are shown in Fig. 7i–l. In particular, all OCT images shown in Fig. 7 have 256 Ascans, which corresponds to his lateral scan range of ~4.4 mm. A smaller lateral extent was chosen to unambiguously yield Figs. 7e–h from the tumor. To perform one-class tissue classification, we divided the OCT images into eight non-overlapping patches (32 Ascans per patch). For every image patch, we followed the procedure shown in Fig. 4a to establish feature vectors for various image patches and output prediction scores using a pre-trained one-class SVM classifier. A positive prediction score corresponded to normal skin tissue and a negative prediction score to abnormal skin tissue. We averaged the scores using the results of all eight patches in the image and summarized the results in Table 2. predicted the tissues to be normal and abnormal, respectively. The tissue classification was correct. On the other hand, scans performed outside the circle delineated by surgery (scans 5–8 in Fig. 7d, images in Fig. 7i–l) provide margin assessments. According to the one-class SVM classification results, Fig. 7i,j (scan 5 and scan 6 in Fig. 7d) correspond to abnormal skin. So the margin was positive. Figure 7k,l (scans 7 and 8 in Figure 7d) correspond to normal skin. So the margin was negative. To verify the results of margin evaluation, the results of histological examination are shown in Fig. 7m,n. Histology suggests a positive margin, which is consistent with the classification results.

(a) Clinical photographs taken from patients. (B.C.) OCT image obtained from the patient’s forearm. (d) scan pattern used for tumor profiling. (e–h) OCT images acquired according to trajectories 1–4 in Fig. 6d. (i–l) OCT images acquired according to trajectories 5–8 in Fig. 6d. (meters) results of histological examination; (n) Mohs histological document showing histologically positive stage 1 margins.