Firdous Saleheen

Non-invasive mechanical properties estimation of embedded objects using tactile imaging sensor

Non-invasive mechanical property estimation of an embedded object (tumor) can be used in medicine for characterization between malignant and benign lesions. We developed a tactile imaging sensor which is capable of detecting mechanical properties of inclusions. Studies show that stiffness of tumor is a key physiological discerning parameter for malignancy. As our sensor compresses the tumor from the surface, the sensing probe deforms, and the light scatters. This forms the tactile image. Using the features of the image, we can estimate the mechanical properties such as size, depth, and elasticity of the embedded object. To test the performance of the method, a phantom study was performed. Silicone rubber balls were used as embedded objects inside the tissue mimicking substrate made of Polydimethylsiloxane. The average relative errors for size, depth, and elasticity were found to be 67.5%, 48.2%, and 69.1%, respectively. To test the feasibility of the sensor in estimating the elasticity of tumor, a pilot clinical study was performed on twenty breast cancer patients. The estimated elasticity was correlated with the biopsy results. Preliminary results show that the sensitivity of 67% and the specificity of 91.7% for elasticity. Results from the clinical study suggest that the tactile imaging sensor may be used as a tumor malignancy characterization tool.

Characterization of Mammary Tumors Using Noninvasive Tactile and Hyperspectral Sensors

The use of both tactile and hyperspectral imaging sensors, which exploit the mechanical and physiological changes in tissues, can significantly increase the performance in automatic identification of tumors with malignant histopathology. Tactile imaging measures the elastic modulus of a tumor, whereas hyperspectral imaging detects important biochemical markers. Spontaneous mammary tumors in canines were used to demonstrate the efficacy of our approach. The tactile sensor achieved a sensitivity of 50% and a specificity of 100% in identifying malignant tumors. The sensitivity and specificity of the hyperspectral sensor were 71% and 76%, respectively. We investigated several machine learning techniques for fusing the tactile and spectral data, which increased the sensitivity and specificity to 86% and 97%, respectively. Our tactile and hyperspectral imaging sensors are noninvasive and harmless (no ionized radiation is used). These imaging sensors may not only eliminate unnecessary surgeries, but will also motivate the development of similar sensors for human clinical use, due to the fact that canine and human tumors have similar physiology and biology.

Tactile Imaging System for inclusion size and stiffness characterization

We developed Tactile Imaging System (TIS), which measures mechanical properties of tissue inclusions. The functional components of TIS include: a monochrome camera, a soft and transparent silicone probe, and an LED illumination circuit. The deformation of the soft TIS probe is used to evaluate size and compliance of inclusions, which is the inverse of elastic modulus. TIS algorithm development and testing were performed using a fabricated silicone tissue phantom with soft inclusions. We completed validation experiments for size and compliance estimation with TIS using the phantom. The results from these experiments suggest that TIS can be used for characterization of tissue inclusions.

Dynamic positioning sensing system for estimating size and depth of embedded object

In this study, we developed a dynamic positioning sensing (DPS) system for estimating size and depth of embedded objects. The DPS System is comprised of a tactile imaging sensor (TIS), a near infrared diffuse optical imaging (DOI) unit. The TIS have a flexible and transparent sensing probe, a LED (light emitting diode) module, and a CCD camera with lens. The light from the LED-illuminated probe is scattered when a force is applied to the probe, and captured by the camera as a tactile image. TIS determined the tumor size by correlating applied force, depth and number of pixel in tactile image. On the other hand, using a near infrared laser as source and the CCD camera as detector, we obtained the diffuse optical images. From these images, we computed the absorption coefficient of the embedded tumor phantom. We maneuvered the source-detector simultaneously for collecting diffuse optics information. We termed this maneuver as dynamic positioning. The TIS provided a priori location information. The combination of the absorption coefficient, tactile data, and dynamic positioning method improved the size and depth estimation. The experimental results showed that the TIS estimated the tumor phantom size with 7.23% error, while the DPS System measured the size with 0.8% error. The TIS depth estimation error was 41.83%, and the DPS System reduced depth measurement error to 20.67%.

Classification of breast masses using Tactile Imaging System and machine learning algorithms

In this study, we used Tactile Imaging System (TIS) and machine learning algorithms to classify breast masses in vivo as malignant or benign. When the silicone probe at the front end of TIS is compressed against the breast mass, the indentation profile of this waveguide is captured by a CCD camera. Then TIS algorithm determines the size and stiffness of inclusions based on the acquired tactile images. The size and stiffness results are then used as the input features for breast tumor classification algorithms. We compared three tumor classification algorithms: k-nearest neighbor, support vector machine, and Naïve Bayes, which are known to work well for limited data set. We tested these algorithms on twelve human breast tumors. The results were evaluated using the leave-one-out cross validation technique. Among the three algorithms, k-nearest neighbor classifier performed the best with sensitivity of 86% and specificity of 100%.

Dynamic imaging system for mechanical and spectral properties estimation

Mechanical and spectral properties of tumor can be used in detecting breast cancer and characterizing the malignancy of tumor. In this study, we develop a Dynamic Imaging System (DIS) for estimating the mechanical (size, depth, and elastic modulus) and spectral (absorption coefficient) properties of embedded objects. The dynamic imaging system consists of a near infra-red light source with a Tactile Imaging Sensor (TIS) as detector, a source-detector maneuvering system, and a laptop as optical data acquisition unit. The source-detector geometry is controlled for collecting diffuse optical information, which we termed as dynamic imaging. This system determines absorption coefficient, and along with tactile information produces improved depth, and elastic modulus estimation. We performed experiments with a multimodal phantom for determining mechanical properties with the dynamic imaging system. The DIS estimated size and depth with errors of 0.8% and 20.67% compared to the tactile imaging sensor measurement errors of 7.23% and 41.83%. The dynamic imaging system also determined elastic modulus and absorption coefficient of the embedded object with moderate errors.

Tactile and hyperspectral imaging sensors for mammary tumor characterization

In this paper, we developed and tested both a tactile as well as a hyperspectral imaging sensor, which exploit the physiological and mechanical changes that occur in malignant tumors. The use of both modalities (tactile and hyperspectral) increases the accuracy of identifying malignant tumors. Spontaneous mammary tumors in dogs were used to test our sensors. The sensitivity and specificity of the fused tactile and hyperspectral data is 67% and 83%, respectively. These imaging sensors will not only decrease the need for unnecessary surgery, but it will also facilitate and jump-start the development of a tactile and hyperspectral imaging sensor for human clinical use because of the similarities between human and canine breast cancer.

Adaptive Neural replication and resilient control despite malicious attacks

In this paper, an Adaptive Neural Control (ANC) architecture is used for system replication and control within a Resilient Control framework. A dynamic model is chosen for our plant and a “maliciously attacked” plant. A Model Reference Adaptive Control (MRAC) architecture is used to replicate and control the plant to match an ideal reference system. At certain time, we replicate a malicious attack by changing plant parameters, injecting false data, or altering sensor data. This attacked plant is then replicated and controlled to match the reference system. Simulations were carried out to show that accurate system replication and resilient control is possible using adaptive neural networks.

Itchy skin region detection using hyperspectral imaging

Itch is the primary symptom of inflammatory skin diseases such as atopic eczema and psoriasis, chronic renal failure, and chronic hepatic failure. Itch, like pain, is a subjective symptom. Characterizing itchy skin and skin prone to itch will lead to better understanding of these symptoms and ultimately better diagnosis and treatment of the underlining disease. The goal of our study is to determine whether the itchy skin region can be detected by hyperspectral imaging. We used an imaging system equipped with liquid crystal tunable filter for collecting hyperspectral images. A halogen lamp was used to illuminate the region of interest. Images were taken from 650 nm to 1100 nm wavelength with 10 nm interval. The hyperspectral images were collected from the forearms of two male and two female subjects. An approximate 50 mm × 50 mm region of interest was marked on the forearms before imaging. The itch was mechanically induced. Imaging was performed for three conditions with a 99% Spectralon white diffuse reflectance target on the side: before inducing itch (normal region), after inducing itch (test region), and after removing itch (control region). Two methods were used to detect the itchy and nonitchy regions from the normalized hyperspectral data. The first method used a spectral distribution exploration method. The second method used a supervised classification method, more specifically, a support vector machine (SVM) algorithm. The spectral distribution exploration method did not detect any different spectral signature for itchy region. On the other hand, the SVM classifier detected the itchy region with the surrounding non-itchy region. These results demonstrated the feasibility of using hyperspectral imaging combined with classification algorithms for detecting itchy skin region.

Evaluating color performance of whole-slide imaging devices by multispectral-imaging of biological tissues

The color reproducibility of two whole-slide imaging (WSI) devices was evaluated with biological tissue slides. Three tissue slides (human colon, skin, and kidney) were used to test a modern and a legacy WSI devices. The color truth of the tissue slides was obtained using a multispectral imaging system. The output WSI images were compared with the color truth to calculate the color difference for each pixel. A psychophysical experiment was also conducted to measure the perceptual color reproducibility (PCR) of the same slides with four subjects. The experiment results show that the mean color differences of the modern, legacy, and monochrome WSI devices are 10.94±4.19, 22.35±8.99, and 42.74±2.96 ▵E00, while their mean PCRs are 70.35±7.64%, 23.06±14.68%, and 0.91±1.01%, respectively.