Convolution Neural Network based segmentation
Our lab recently validated a radiomic pipeline which allows us to predict with considerable accuracy whether a woman with breast cancer has suffered metastasis to the sentinel lymph nodes.
Whereas the clinical adoption of such a pipeline could prevent thousands of women from having unnecessary, invasive procedures done every year, our pipeline (as with generally all radiomic analyses) require someone to segment out lesions. This is an extremely time consuming process and is a roadblock for clinical radiomics.
Recently, convolutional neural networks have emerged as a dominant technique for virtually any image recognition or processing task. We have become interested in the ability of such models to accurately reflect the radiologist-drawn ROIs that we currently feed into our model.
With only very little training data (106 scans each comprising only about 3-10 slices containing a lesion) we have established a convolutional neural network that allows us to segment breast lesions with sufficient accuracy to input into our radiomics pipeline without impairing our results.
CNN based transmission image synthesis for attenuation correction
PET/MRI has been considered an ideal environment for neuroimaging-driven research; such research is integral in diverse fields such as neurology, psychiatry and oncology. Despite the many expected benefits of PET/MRI, challenges in attenuation correction remain a limiting factor in the quality of PET data collected with this new modality.
Currently, techniques are being developed to try to "predict" patient-specific CT data from their MRI scans, which allows researchers to correct for attenuation effects. That said, CT is not an ideal standard for attenuation correction as CT photons are ~5x lower energy than PET photons, which means they have different attenuation properties.
Recently, our lab has implemented a novel approach to this problem by generating PET transmission data--which are acquired with the same energy photons as PET--from MRI only by using a convolutional neural network. By basing our network off of higher fidelity attenuation data, we believe that we can create a more optimal pipeline for processing PET/MRI brain imaging data.
This neural network approach also exhibits a few improvements over the methodology used for other currently popular algorithms. Most notably, our method doesn't require input MRI data to be co-registered onto our input data, as CNNs are pretty good at dealing with translations by design.
We've recently validated our approach using [11C]-WAY and demonstrated that our method can reliably reproduce both static and dynamic PET data. The latter fact is crucial for psychiatric research in particular, where researchers often which to quantify neurobiological interests such as neuroreceptor density, which requires sophisticated kinetic modelling.