The French company was created in 2021 with the aim of developing software solutions to help nuclear physicians in their daily routine. It all started with this software as medical device that detects, segments and quantifies abnormalities on PET-CT FDG examinations. Augmenting nuclear physicians with such a sofware enables them to save time, avoid errors and access complex biomarkers in clinical routine.

This software has been built on the largest database of annotated PET-CT FDG examinations. It puts great effort to make this database heterogenous and representative of each manufacturer, modality type, modality age, various type of configuration pathologies (indicated for FDG) and patient age (above 18).

It is intended for assisting nuclear medicine physicians in the detection and analysis of hypermetabolic foci in PET-CT examinations, after their review of images and before the writing of the report.

It acts frictionless for the physician. Such a software triggers automatically the computation and image generation when PET-CT series are received on the DICOM node. These series are either sent from modalities automatically (by adding a routing rule to the protocol) or by manually sending exams from a viewer or PACS through the DICOM messaging protocol.

The two only required series are an attenuation corrected PET and a CT DICOM series.

The software outputs DICOMS interpretable by physicians in their workstation or PACS systems. Output DICOM format include among others secondary captures DICOM files or DICOM RT-struct files.

The detection algorithm leverages recent advances in informatics and mathematics called Deep Learning. When first machine learning techniques used to require engineers to develop hand-made patterns to be detected (spheres, squares, sharp intensity difference), the Deep Learning enable to automatically learn the patterns to be detected within an image by providing it with several thousands of manually annotated examinations, for hundreds of time. The result called a “trained neural network” is then rigorously evaluated on hundreds of exams by professional nuclear physicians to assess its performance and benefit.

The French company has trained an AI to automatically detect lesions on the basis of tens of thousands of lesions manually annotated by nuclear medicine experts.

It shall be noted that the software receives a copy of the examination images and works only on the two series PT AC and CT. It summarizes its findings in the form of “Secondary Capture” (which can be assimilated to screenshots), which are sent to the physician s viewer and comes in addition of all the examination series: no image update can be performed.

Within optimal conditions, the device can help nuclear medicine physicians to localize abnormal FDG uptake. The use of this medical device in clinical routine brings several benefits to the patient and clinician. Clinical benefits for the patient include but are not limited to a more accurate staging, better evaluation of the pathology. Clinical benefits for the clinician include but are not limited to faster and more accurate image analysis.

The French company is currently looking for clinical partnerships to validate its solution and continue to bring high-value-added functionalities to clinicians.
Advantages and innovation
Today, oncology diagnosis is too prone to variability (inter-physician, intra-physician, inter-geographical, etc.) because it is based on multiple, non-exhaustive, hand-measured indicators. The aim of this innovative medical device is to create a platform to gather robust, quantitative, reproducible and exhaustive indicators for routine clinical use, in order to improve the management and follow-up of pathologies over time. The innovation of this medical device lies in its technological approach. While it uses deep-learning in medical imaging, no one has yet focused on PET-CT examinations. PET-CT are extremely complex:

- It is performed on whole body, i.e. from head to toe, and is therefore obviously more voluminous than a localized examination of a single organ.
- It is 3-dimensional, which considerably increases the size of the examination.
- It is hybrid, i.e. it is a PET scan and a CT scan, i.e. it contains twice as much data.

These technical complexities are compounded by the complexity of annotation: as these exams are very voluminous and cover the entire patient, it takes almost 1 hour of doctor time to annotate them. This represents a considerable cost. Since its launch, this medical device has developed methods to reduce this time considerably, enabling large quantities of data to be taken into account.

This determination to start with the most complete imaging examination is therefore quite unique, and goes against the grain of what is being done in the world of AI in medical imaging. It s a gamble that has paid off, because the marketing and good performance of its first medical device shows that it is capable of meeting these technical challenges.

This medical device is the first complete algorithm based on deep learning, giving it first-rate performance.
Technical Specification or Expertise Sought
This software has to be installed by the French company s team on an Ubuntu 22.04 (Desktop or Server), with at least 4 CPUs and 16Go RAM, and a storage capacity of at least 100Go.

IT network shall enable a fixed IP to be provided to the server, which is required in the DICOM standard for DIMSE protocol; it shall also enable connections between images senders (like viewers, pacs and modalities) to the server on the corresponding port.

The software is intended for use in healthcare facilities such as hospitals, clinics and radiology centers, including tele-radiology centers.

The user will use the software s output images with the same viewer he / she is used to: no specific work environment is required apart from their usual ones.

The software is intended to be installed within a DICOM nodes network as it leverages the corresponding interoperability standard DICOM. Thus, the software is compatible with all common imaging systems, viewers, and PACS.

The software is intended to be installed on a dedicated server (physical or virtual machine). The server s security is under the responsibility of the customer and does not require special treatment beyond being stored in a trusted environment / room.

Based on the device performance tests and verification and validation activities, the following performance is claimed by the French company for its medical device:
- Automatic visualization and labeling of high uptakes foci with a minimum of Dice Similarity Coefficient more than 0.6 on average
- Sensitivity of at least 90% for all lesions detection independently of lesion uptake
- 85% of agreement at the liver PERCIST threshold of described lesions in medical reports
- more than 90% of spearman correlation for TMTV at baseline for DLBCL