Ultra-fast/Low-dose PET/CT imaging Using Transformers
Mohammad Mashayekhi, Amirhossein Sanaat, Narges Aghakhan Olia, Zahra Khazaei, Arian Amiramjadi, Habib Zaidi
This work aims to explore the performance of novel deep learning approaches in PET administrated dose reduction. Dose reduction in nuclear medicine modalities would lead to image quality degradation and low diagnostic value. This study enrolled 76 clinical PET images from a Biograph mCT PET/CT scanner acquired in the list-mode format. Retrospectively collected data were used to propose a low-dose(LD) to full-dose(FD) translation model. All patients underwent the standard scan protocol with a 20 min acquisition time. 5% of the collected List-mode events were utilized to simulate the corresponding LD counterparts. A vision transformer network comprised of transformer and convolutional blocks were implemented to predict FD data from the LD images in the image space. The quality of synthetic PET images generated by the transformer model was assessed by some standard quantitative metrics like peak signal-to-noise ratio (PSNR), root means squared error (RMSE), and structural similarity index measure (SSIM), and Correlation parameter. Moreover, a joint histogram analysis was performed on the LD and predicted FD images with respect to the reference FD images. Regarding the quantitative analysis, the SSIM, PSNR, and Correlation rose considerably by 25%, 87%, and 7.6%, respectively, as opposed to an 80% decline for RMSE. The joint histogram plotted for the generated FD images resulted in a superior agreement with standard FD data in comparison with the LD data. Altogether, based on the visual illustration, the noise was suppressed effectively and the underlying information was preserved appropriately in the predicted FD images.
A machine learning web application for screening social anxiety disorder based on participants' emotion regulation (ML-SAD)
Sara Ahmadi Majd, Mohamad Rasoul Parsaeian, Mohsen Madani, Manouchehr Moradisabzevar, Abolfazl Mohammadi
Social Anxiety Disorder (SAD) is called a neglected anxiety disorder since people do not realize its existence and the need to be evaluated by an expert. Thus, it is important to develop widely available self-screening systems to assess people and guide needy people for further evaluations. Consequently, in this paper, we present a machine learning-based web application to screen SAD. The web application consists of 10 multimedia scenarios with which people with SAD may have difficulty dealing. Four hundred eighty-eight subjects were asked to consider themselves in these scenarios and rank their competency in dealing with each specific situation considering three emotion regulation strategies. The participants were divided into two groups, SAD and No SAD, based on their diagnostic history of SAD and their self-assessment of their anxiety level. Then, a random forest was trained and tested to screen the subjects with SAD from the No SAD subjects with over 85% accuracy.
Radious: Unveiling the Enigma of Dental Radiology with BEIT Adaptor and Mask2Former in Semantic Segmentation
Mohammad Mashayekhi, Sara Ahmadi Majd, Arian Amiramjadi, Babak Mashayekhi
X-ray images are the first steps for diagnosing and further treating dental problems. So, early diagnosis prevents the development and increase of oral and dental diseases. In this paper, we developed a semantic segmentation algorithm based on BEIT adaptor and Mask2Former to detect and identify teeth, roots, and multiple dental diseases and abnormalities such as pulp chamber, restoration, endodontics, crown, decay, pin, composite, bridge, pulpitis, orthodontics, radicular cyst, periapical cyst, cyst, implant, and bone graft material in panoramic, periapical, and bitewing X-ray images. We compared the result of our algorithm to two state-of-the-art algorithms in image segmentation named: Deeplabv3 and Segformer on our own data set. We discovered that Radious outperformed those algorithms by increasing the mIoU scores by 9% and 33% in Deeplabv3+ and Segformer, respectively.
L2 Norm Guided Adaptive Computation
Mani Shemiranifar, Mostafa Dehghani
Although the human brain can adjust the amount of time and energy it uses to solve problems of varying complexity, many standard neural networks require a fixed computation budget regardless of the problem’s complexity. This work introduces L2 Adaptive Computation (LAC), a new algorithm that adjusts the computation budget, by tracking changes in the L2 norm of a neural network’s hidden state as layers are applied to the input. Unlike previous methods, LAC does not require additional trainable modules or auxiliary loss terms to make halting decisions. LAC matches the results of best-performing methods on a complex synthetic task and improves image classification accuracy while also increasing efficiency.