In nuclear Medicine, radioactive compounds are administered to patients in order to get both functional and anatomical data. With Anger Cameras and specific cristals, one can collect an image from the distribution of the tracer, which is called a “scintigraphy”.
In thyroid scintigraphy, several pathological situations may occur. François-Xavier Hanin, nuclear medicine physician (CHU UCL Namur) developed a deep learning model based on a U-net architecture to automatically analyze these images. First, a segmentation system was developed to define the contours of the thyroid. This has potential applications in automatic assessment of the uptake, a computation of the percentage of injected activity which is absorbed by the thyroid.
He then used a similar architecture to detect cold nodules. These nodules have a particular interest in thyroid pathology, as many differentiated thyroid cancers appear as cold nodules (meaning, with no uptake).
The U-net Architecture was described in 2015 (1) and its implementation was applied with 6 levels, a maximum of 1024 filter dimension at center, 2 convolutions on the way down and 3 on the way up, Batchnormalization after every convolution, and concatenation with corresponding levels on the up branch.
Training was performed using a database of 1342 thyroid scintigraphies. Data augmentation by horizontal flip only allow the use of 2684 images, with corresponding masks drawn by the same operator. Training used 80 % of the database images for 150 epochs. Mean Dice coefficients were 0,98 ± 0,01 on test images for segmentation. The cold nodule detection showed an accuracy of 96.5 compared to a human trained reader.
The automatic contour delineation has a potential impact on clinical wokflow for several reasons. First, it allows an automated segmentation for the uptake calculation. Second, it opens a wide range of radiomics analysis. The cold nodule detector brings an added value to the clinician for small lesions.
His current work has many more fields of interest in nuclear medicine and radiation oncology:
-Automatic assessment of uptake and its validation
-Implementation in clinical routine of segmentation and cold nodule detection
-Implementation of radiomic features after segmentation
-Detection of suspected metastatic hot spots in bone scintigraphy
-Automatic delineation of at-risk organs in radiation oncology (CT-based)
(1) U-Net: Convolutional Networks for Biomedical Image Segmentation – https://arxiv.org/abs/1505.04597
https://www.linkedin.com/pulse/deep-learning-nuclear-medicine-practical-example-how-hanin/