12.12.2020 Skin cancer research with convolutional neural networks and hyperspectral camera (Annala)
Opponent Professor Keijo Ruotsalainen (University of Oulu) and Custos Docent Illka Pölönen (Ä¢¹½Ö±²¥). The doctoral dissertation is held in Finnish.
The audience can follow the dissertation online.
Link to the Zoom Webinar (Zoom application or Google Chrome web browser recommended):
Phone number to which the audience can present possible additional questions at the end of the event (to the custos):
Leevi Annala's dissertation on 12th of December 2020 at 10.00 considers the analysis of hyperspectral images using the convolutional neural network and stochastic simulation.
These methods can be used to study for example skin cancer. Hyperspectral images taken of skin cancer can be used, for example, to distinguish between diseased and healthy skin, to define the boundaries of skin changes and to classify different types of skin cancer. Stochastic simulation, in turn, can produce real-looking hyperspectral data that can be used to train various machine learning algorithms, such as the already mentioned convolutional neural network, and thereby to more accurately identify skin cancer from hyperspectral images.
Hyperspectral imaging combines a device that transmits only light of a certain frequency and a standard machine vision camera. This allows one to adjust the wavelength of the light entering the camera's cell, and taking multiple images with adjusting the wavelength entering the cell between them. After the imaging process, there are about 50 to 200 photographs in a pile, all corresponding to some small wavelength range of light. Such a method provides more accurate information about the color of the subject compared to the human eye or an ordinary camera.
The convolutional neural network is especially known for pattern recognition tasks. The convolutional neural network learns to identify suitable shapes, characters, textures and other features from images/data. This information can be combined with identifiable items. For example, if the picture has the shape of a car, and a metallic texture, it’s probably a photo of the car. If, on the other hand, the texture is like paper, it may be a drawing. Hyperspectral data can be searched for both image-oriented features and spectral-oriented features using the convolutional neural network.
The stochastic simulation performed in the dissertation is based on the Markov process. The Markov process can be modeled as a state diagram, where there is some probability associated with the transition from one state to another. In the case of hyperspectral data and, for example, skin cancer images, these probabilities can be calculated based on the physical properties of the skin, giving a fairly accurate model of how light travels in different layers of the skin and how much light is absorbed into the skin at different wavelengths.
Other applications of the studied methods are, for example, in quality control in the paper and pharmaceutical industries and in agriculture and forestry.