Computer-Aided Diagnosis Algorithm for Classification of Malignant Melanoma Using Deep Neural Networks.

Malignant melanoma accounts for about 1-3% of all malignancies in the West, especially in the United States. More than 9000 people die each year. In general, it is difficult to characterize a skin lesion from a photograph. In this paper, we propose a deep learning-based computer-aided diagnostic algorithm for the classification of malignant melanoma and benign skin tumors from RGB channel skin images. The proposed deep learning model constitutes a tumor lesion segmentation model and a classification model of malignant melanoma. First, U-Net was used to classify skin lesions in dermoscopy images. We implement an algorithm to classify malignant melanoma and benign tumors using skin lesion images and expert labeling results from convolutional neural networks. The U-Net model achieved a dice similarity coefficient of 81.1% compared to the expert labeling results. The classification accuracy of malignant melanoma reached 80.06%. As a result, the proposed AI algorithm is expected to be utilized as a computer-aided diagnostic algorithm to help early detection of malignant melanoma.

Smartphone-based Compression-Induced Scope with Temperature Sensor for Inflammatory Breast Cancer Screening.

The Smartphone-based Compression-induced Scope (SCIS) is a mobile device designed to sense the mechanical properties of tumors. Here, an SCIS system with an infrared temperature (SCIS-T) sensor is developed. The color and texture information of target skin are extracted from the SCIST images using a color-based edge detection technique and a texture filter. This new system provides mechanical properties (size, elasticity) of the inclusion as well as the skin surface (color, temperature, texture) characteristics. The application of this system is in the identification of inflammatory breast cancer, which is characterized by color, texture, and temperature change. The device is tested using chicken breast phantoms with embedded silicone inclusion.

The tactile sensation imaging system for embedded lesion characterization.

Elasticity is an important indicator of tissue health, with increased stiffness pointing to an increased risk of cancer. We investigated a tissue inclusion characterization method for the application of early breast tumor identification. A tactile sensation imaging system (TSIS) is developed to capture images of the embedded lesions using total internal reflection principle. From tactile images, we developed a novel method to estimate that size, depth, and elasticity of the embedded lesion using 3-D finite-element-model-based forward algorithm, and neural-network-based inversion algorithm are employed. The proposed characterization method was validated by the realistic tissue phantom with inclusions to emulate the tumors. The experimental results showed that, the proposed characterization method estimated the size, depth, and Young's modulus of a tissue inclusion with 6.98%, 7.17%, and 5.07% relative errors, respectively. A pilot clinical study was also performed to characterize the lesion of human breast cancer patients using TSIS.

Topology preserving relaxation labeling for nonrigid point matching.

This paper presents a relaxation labeling process with the newly defined compatibility measure for solving a general nonrigid point matching problem. In the literature, there exists a point matching method using relaxation labeling; however, the compatibility coefficient takes a binary value of zero or one depending on whether a point and a neighbor have corresponding points. Our approach generalizes this relaxation labeling method. The compatibility coefficient takes n-discrete values which measure the correlation between point pairs. In order to improve the speed of the algorithm, we use a diagram of log distance and polar angle bins to compute the correlation. The extensive experiments show that the proposed topology preserving relaxation algorithm significantly improves the matching performance compared to other state-of-the-art point matching algorithms.

Inclusion mechanical property estimation using tactile images, finite element method, and artificial neural network.

In this paper, we developed a methodology for estimating three parameters of tissue inclusion: size, depth, and Young's modulus from the tactile data obtained at the tissue surface with the tactile sensation imaging system. The estimation method consists of the forward algorithm using finite element method, and inversion algorithm using artificial neural network. The forward algorithm is designed to comprehensively predict the tactile data based on the mechanical properties of the tissue inclusion. This forward information is used to develop an inversion algorithm that will be used to extract the size, depth, and Young's modulus of a tissue inclusion from the tactile image. The proposed method is then validated with custom made tissue phantoms with matching elasticities of typical human breast tissues. The experimental results showed that the proposed estimation method estimates the size, depth, and Young's modulus of tissue inclusions with root mean squared errors of 1.25 mm, 2.09 mm, and 28.65 kPa, respectively.

Determining the operative line of resection for image-guided emphysema surgery using a laser scanner and non-rigid registration.

BACKGROUND: Although many diseases such as emphysema are diagnosed with preoperative imaging modalities, this information is rarely utilized in the operating room. A method that relates the preoperative images to the non-rigid organ in physical space would aid a surgeon to determine the line of resection. METHODS: We used a three-dimensional (3D) laser scanner to obtain intraoperative images of the lung and overlayed it with preoperative CT images, using a non-rigid image registration method. RESULTS: The non-overlapping registration error of the system was 1.91 +/- 0.28% without organ deformation and 2.69 +/- 0.28% with 9% organ deformation. When 83% of the organ was visible, the registration error was 2.99 +/- 0.42%. CONCLUSION: A novel image overlay system using a 3D laser scanner and a non-rigid registration method was implemented and its accuracy evaluated. By using the proposed system, we successfully related the preoperative images with an open organ in the operating room.

Characterization of lung tissues using liquid-crystal tunable filter and hyperspectral imaging system.

Hyperspectral imaging system has been developed to characterize lung tissue for detecting emphysematous tissues in lung volume reduction surgery. The system consists of a charge-coupled device and liquid crystal tunable filter, which is continuously tunable in the near-infrared spectral range of 650 - 1100 nm with a mean bandwidth of 5 nm. Using hyperspectral data, the spectral signature of healthy lung tissue and simulated smokers lung tissue is obtained and compared. The data show the peak absorption intensity at four different wavelengths (760, 805, 915, and 970 nm). However, the reflectance intensity of simulated smoker's lung tissue over all spectral range is considerably higher than the normal lung tissue. The differences provide the basis for the detection and characterization of emphysema from healthy lung tissue.

Estimation of operative line of resection using preoperative image and nonrigid registration.

Even though accurate diagnosis of organs is done using preoperative images such as CT or MRI, these information are not directly used in the operating room, because organs are nonrigid and their shapes change with time. In this paper, we propose to obtain an intraoperative image of an open organ and fuse the image with a preoperative image. The intraoperative image is obtained from a three-dimensional laser scanner. The registration of preoperative image to the intraoperative image can relate the information from the preoperative image to the open organ in the operating room. We do this by registering preoperative images to intraoperative images. An algorithm based on Robust Point Matching method is developed for this nonrigid image registration problem. We also propose a new metric called Non Overlapping Ratio to determine the registration error. The experiments demonstrate that the proposed method is capable of achieving region of interest estimation within 1.51 mm mean distance error and 0.66% Non Overlapping Ratio.