Automatic detection of scintillation light splashes using conventional and deep learning methods
Six methods for the automatic detection of scintillation light splashes in a portable gamma camera are compared. Each imaging frame might contain any number of light splashes (including none), and the location and size of each light splash must be identified. For real-time imaging, splashes must be identified and characterised quickly and with minimal processing overhead. The techniques are compared on their ability to accurately determine the number, position, and size of light splashes, and to reconstruct the deposited energy within each splash for a simulated data set with known ground-truths. The speed of each technique and the ease of implementation are also discussed. For accuracy in blob (light splash) localisation, a Laplacian of Gaussian approach was found to provide the most accurate estimation. However, its performance greatly relies on the appropriate tuning of preprocessing parameters prior to image analysis and the number of blobs in each frame. Deep learning approaches (Faster-RCNNs) performed significantly better than traditional algorithms in terms of predicting the size of each light splash, did not require image preprocessing and were also more stable over a range of frame occupancies. Moreover, the paper fine-tuned a VGG16 based Faster-RCNN model with the simulated data set for the scintillation light splash detection, called DeepSplashSpotter (DSS).
STFC Cancer Diagnosis Network+ (ST/S005404/1)
- Computer Science