Lung Cancer Identification Through Artificial Neural Network
Most of the time patients are diagnosed with delay and the tumor is already at an advanced stage. At that time it is often very hard to intervene with a promising cure.
The detection pipeline of this disease takes various steps and the procedures followed may be different. It goes through imaging tests (X-ray, CT, CT/PET), which require careful screening of the results made by radiologists to individuate the tumor mass, classifying the type and determining the stage. In order to choose the most appropriate treatment, it is necessary to have information on the type and stage of the tumour, information that need a highly invasive examination such as biopsy; in fact, it requires surgical procedures (e.g. bronchoscopy, thoracoscopy, mediastinoscopy).
Abnormal cells, if malignant, can spread across different areas of the human body before their detection and create permanent damages. If this happens, treating the disease will be much more difficult.
Since, most of the time, an early detection can be lifesaving, we are willing to provide to the radiologists and the physicians an automated pipeline tool that could reduce the amount of time spent on each medical exam and could rise the accuracy of each result.
The radiologist will still have to interpret the final results, since the algorithm cannot replace the decision-making process of a radiologist but only optimize the whole process.
As we can observe, this is a very resource-intensive process, requiring both expert physicians and medical imaging technologies.
In order to improve this medical area our team is going to focus on the three main goals: the first one is the correct segmentation of the lungs, the second one consists in the identification of the tumor masses and, eventually, the third goal will be their classification, that means distinguish which mass is primary and which is metastatic.
We want to do so by applying Machine Learning algorithms to the analysis of Computed Tomography, as, for this purpose, it is the most useful diagnostic test.
As first step, we want to segment each CT exam to identify the lungs, erasing all the non-informative structures in the images.
Secondly, we will localize the tumor masses, defining their contours on each image so that the result can be easily read and understood from the radiologists.
As a possible final step, once the mass has been identified, through a radiomic-based feature extraction we would collect some interesting insights concerning the tumor mass.
Considering the problems mentioned in the project summary, from the long time taken by radiologists in the segmentation of images to the high invasiveness of examinations, we think that our project can improve the diagnosis of lung cancer.
In order to train the neural network and obtain the most accurate results, it would be extremely useful for us to have a dataset of CT images of both healthy and sick patients.
The datasets that NLST makes available would suit perfectly our intentions and would help us in the development of the work path chosen in order to accomplish our goals.
Marco Domenico Santambrogio:
Associate Professor at Politecnico di Milano
marco.santambrogio@polimi.it
Eleonora D’Arnese:
Ph.D. student in Information Technology, Computer Science
eleonora.darnese@polimi.it
Sara Caramaschi:
bachelor student of Biomedical Engineering, Politecnico di Milano, caramaschisara@gmail.com.
Irene Canavesi:
bachelor student of Biomedical Engineering, Politecnico di Milano,
irene.canavesi@mail.polimi.it