Pulmonary Nodule Detection by a Team of Medical Practitioners
Industry: Medical Software
Size: 50-100 employees
Our client provides artificial intelligence software for full body imaging. With the help of deep learning and AI algorithms, this software analyzes medical images and patient data to help radiologists detect critical anomalies.
To create advanced healthcare-grade AI-based decision support software, massive amounts of unsorted and encrypted lung scans needed to be annotated. These scans had to be sorted, grouped, and checked for pulmonary nodules. Pulmonary nodules then had to be marked as prominent, non-prominent, borderline, or granuloma.
Target Time for Each CT scan
Volume of CT scans
Amount of Nodules to be Detected
Our main challenge was that no specific instructions nor any annotation tools were provided to our team. The output format was not standard, so the conversion took more time and effort than anticipated.
Mindy Support worked with three radiologists each who had 5+ years of experience and were also members of the Radiologist’s Association of Ukraine, and one was also a member of the European Society of Radiology. A dedicated project manager was also provided.
We sorted studies and matched them with their descriptions.
We researched different annotation tools, selected the ones which suited our needs, and set them up.
Radiologists labelled pulmonary nodules in the CT scans with polygons and added special tags for each type of nodule: prominent, non-prominent, borderline, and granuloma. If nodules could not be detected, image artifacts were applied: breathing movement, lung disease, metastatic disease.
For our client’s convenience, we sent the data in two separate formats — JSON and PKL. In order to convert JSON into PKL, we developed a special script.
Our dedicated team of 3 radiologists and 1 project manager delivered the following results:
10k+ nodules detected
99% annotation accuracy
65% cost efficiency
Target time for each CT
nodules to be detected