AI in Molecular Imaging

Introduction

The speaker introduces themselves and their affiliation, and expresses excitement about the potential of AI in medical imaging.

• The speaker works for Siemens Medical Solutions and is part of the PET physics and reconstruction group in Knoxville, Tennessee.

• AI has produced state-of-the-art results across many disciplines, including medical imaging.

• Siemens has invested a tremendous number of resources into the AI space, resulting in a large assortment of AI-driven applications that are already available on the product.

Siemens' Investment in AI

The speaker discusses Siemens' investment in AI and how it has helped procure an incredibly large repository of Medical Imaging data.

• Siemens is one of the world leaders in healthcare-related AI patents.

• They have invested a tremendous number of resources into the AI space, resulting in a large assortment of AI-driven applications that are already available on the product.

• They have 200 dedicated scientists across multiple sites working closely with clinical experts to identify real medical problems and come up with creative ways to address them with AI.

• This effort has helped procure an incredibly large repository of Medical Imaging data.

Explosion of Interest in AI

The speaker shows a plot representing the explosion in interest in AI over recent years.

• There has been an explosion in interest in AI over recent years due to several factors, including advances in hardware and software making these tools more accessible.

• Even now, it would be hard to find a field that hasn't been affected or transformed by Ai.

• Medical Imaging and imagery construction is no exception.

Current Applications at Siemens

The speaker discusses some current applications at Siemens that use artificial intelligence.

• Flow motion uses an intelligent algorithm for organ segmentation to tailor continuous bed motion scans specifically to the patient and improve patient workflow.

• Oncofreeze AI was introduced to address respiratory motion within PET reconstruction, using AI to identify the optimal space within the emission data to extract the respiratory waveform for use in motion correction.

• Multi-parametric PET AI uses artificial intelligence to automatically segment regions in arterial blood so that image-derived blood input functions can be extracted for use in compartmental and fully quantitative parametric analysis.

• Auto ID is a smart algorithm for automatically identifying and segmenting areas of abnormal uptake in PET images while differentiating them from normal physiological uptake.

Current Work at Siemens

The speaker introduces current work being done at Siemens related to AI.

• Siemens' efforts can be classified into three general categories: system hardware, reconstruction, and post-reconstruction.

• Within the hardware bubble, they are investigating ways to use AI to improve the speed of QC and system calibration protocols. They are also looking into using artificial intelligence for improving detector performance itself.

AI in Molecular Imaging

In this section, the speaker discusses the use of AI in molecular imaging and its importance. They also highlight the need to evaluate these algorithms carefully and consistently with clinical needs.

Importance of AI in Molecular Imaging

• The speaker explains that AI is used because it allows us to solve problems that we were unable to before, and it's a very flexible framework.

• AI algorithms have demonstrated impressive robustness to noise in data, which is important for molecular imaging data.

• The speaker suggests using deep learning for sparse count data or reconstructing frames of a dynamic protocol.

Evaluating AI Algorithms

• The speaker presents an example where direct deep learning reconstruction can generate an image faster than iterative algorithms. However, we need to ensure that we preserve quantification across the entire field of view.

• We know that we can suppress high-frequency noise in images with AI denoising. Still, we must consider the clinical implications of this and whether it improves signal detectability lesion SNR.

• We cannot discount the possibility that sometimes suppressing lesion signals can lead to hallucinating or boosting up artificial structures that aren't there. This could inflate confidence levels and lead to incorrect decisions.

Synthesizing Synthetic CT or New Maps

• The speaker mentions synthesizing synthetic CT or new maps from uncorrected emission data but questions whether these emission data reflect quality in a traditional sense.

Investigating the Use of Time-of-Flight PET for Accurate Quantification

The speaker discusses the use of time-of-flight PET to improve signal-to-noise ratio and emission measurements, as well as enhance corrections such as attenuation. They also explore the potential of using convolutional neural networks to achieve accurate quantification and motion correction in PET reconstruction.

Time-of-Flight PET

• Applying time-of-flight PET improves signal-to-noise ratio and emission measurements.

• Histo images provide a representation of acquisition data directly in the same space as target image, making it a straightforward problem for convolutional neural networks.

• Investigating with collaborators at University of Pennsylvania to achieve accurate quantification and incorporate motion correction in PET reconstruction.

Attenuation Correction

• LSO detectors emit background radiation that can be used to identify events from positron annihilation in crystal.

• Transmission measurement even with tracer on board can serve as a good prior for an MLACF or MLA algorithm, but still experiences crosstalk effects in certain places of images.

• Using transmission measurement as prior for an MLACF algorithm and training against real CT or mu map data can improve quality.

Overall, the speaker discusses how time-of-flight PET can be used to improve signal-to-noise ratio and emission measurements, while also enhancing corrections such as attenuation. They explore the potential of using convolutional neural networks for accurate quantification and motion correction in PET reconstruction. Additionally, they discuss how LSO detectors emit background radiation that can be used for attenuation correction, but still experience crosstalk effects. Finally, they suggest using transmission measurement as a prior for an MLACF algorithm trained against real CT or mu map data to improve quality.

Attenuation Correction Improvements

The speaker discusses improvements to attenuation correction in PET CT imaging, specifically investigating the use of deep learning for deformable intermodal re-registration.

Investigating Deep Learning for Deformable Intermodal Re-Registration

• Attenuation correction is acquired not just for PET attenuation correction but also serves a diagnostic purpose.

• Investigating deep learning for providing deformable intermodal re-registration.

• Using convolutional networks to extract feature maps from CT and PET images and concatenate them in a downstream neural network to regress deformation or motion between them.

• Motion field can be used to resample the CT to match the PET distribution.

• Investigating this method for whole-body imaging, estimating respiratory motion, bulk movement of head, arms, legs and feet.

Other Potential Applications

• Potentially applying this technology towards slightly different applications such as using a single CT and applying it to every phase of a respiratory gated series or matching attenuation for consecutive scans on different days.

• Investigating intramodal registration and estimation of motion vectors for incorporating motion correction within the PET reconstruction itself.

Conclusion

• Focused on AI as a tool but also focused on evaluation in clinical context with input from clinical and academic colleagues.

• Acknowledges collaborators who contributed material presented in the talk.

Thank You Note

In this section, the speaker expresses gratitude to Joe for something.

Expressing Gratitude

• The speaker thanks Joe and expresses appreciation.