An intelligent system for lung cancer diagnostics

Researches from Peter the Great St.Petersburg Polytechnic University (SPbPU) in collaboration with radiologists from St.Petersburg Clinical Research for Specialized Types of Medical Care (Oncological) have developed an intelligent software system for lung cancer diagnostics. This software, which can be installed on any computer, analyzes results of patients’ computed tomography (CT) in 20 seconds and provides an image in which the tumor is clearly marked. Researchers have named the system Doctor AIzimov (AI for Artificial Intelligence) in honour of the science-fiction writer Isaac Asimov, who developed three famous laws of robotics.

---

At the end of 2018, the first tests of the system were carried out. The system analyzed anonymized CT images of 60 patients of the Oncological Center. According to the radiologists, the tests were successful, as the system has detected even small-sized focal nodules in lungs (2 mm). “Initially, we set the algorithm to search for nodules starting from 6 millimeters, because it follows the technique applied by radiologists. But the system is so smart that it was able to find nodules of even smaller size”, said the project leader Lev Utkin, the head of the SPbPU Research Laboratory of Neural Network Technologies and Artificial Intelligence.

Research team includes the staff of the University (Lev Utkin, Mikhail Ryabinin, and Alexei Lukashin), experts from the St. Petersburg Oncological Center (the head of the Radiology Department Anna Meldo and a radiologist Ivan Prokhorov). The project was supported by the Russian Science Foundation.

The proposed and developed approach to the lung cancer classification is based on the method of chords: points are randomly drawn on the surface of a nodule in CT image, which are then connected by lines (chords). The histogram of lengths of the chords reflects the shape and the structure of the tumor. Thus the system examines every nodule from the inside, although its external surroundings have significant influence. To take this influence into account, the tumor is virtually placed in a cube, and perpendiculars are drawn from the edges of the cube to the surface of the nodule. Hence, instead of classifying a graphically complex and heavy images of the CT (the size of every CT image is approximately 1 GB), the nodule is represented in the form of compact and simple histograms, which are then analyzed by the Doctor AIzimov system.

The scientists have also trained the system to distinguish malignant and benign tumors. “Many different objects may be detected on the CT images, so the main task was to train the system to recognize what each of the objects represents. Using the clinical and radiological classification, we are trying to train the system not only to detect tumors, but also to distinguish other diseases similar to cancer,” comments Anna Meldo, the head of the Radiology Department of the St. Petersburg Clinical Research Center for Specialized Types of Medical Care (Oncological). The system was trained by analyzing 1000 CT images from LUNA 16 and LIDC datasets. Russian researchers have also collected their own dataset named LIRA – Lung Intelligence Resource Annotated. Currently, the dataset holds CT images of about 250 patients. The scientists are planning to increase the number of images by four times by the mid-2019.

The open testing of the intelligent system will be carried out in 2019. The system will be at first used at the St. Petersburg Clinical Research Center for Specialized Types of Medical Care (Oncological). In the future, the project will be extended and more medical institutions will be involved into the intelligent CT image processing. The system will be adapted to analyze the results of the ultrasound and X-ray medical investigation of other organs. All data will be processed by the supercomputer, and the results issued by the system will be sent to doctors for them to make a decision about the treatment.