Bilkent University
Department of Computer Engineering
M.S.THESIS PRESENTATION
CMCL-NET: CONTRASTIVE MULTI-CHANNEL LESION NETWORK FOR LOW-DOSE PET DENOISING
Emir Türkölmez
Master Student
(Supervisor: Prof.Dr.Selim Aksoy)
Computer Engineering Department
Bilkent University
Abstract: Positron emission tomography (PET) is a nuclear medicine imaging method that reflects in-vivo radiotracer distribution. Reducing injected activity or acquisition time is clinically desirable, but it also makes low-dose PET images noisier and may distort diagnostically relevant uptake patterns. Motivated by obtaining image quality close to standard-dose without additional radiation exposure, this thesis proposes Contrastive Multi-Channel Lesion Network (CMCL-Net) for low-dose PET restoration. CMCL-Net follows an encoder-decoder design with skip connections and uses residual prediction for PET restoration. For a stack of neighboring low-dose input axial slices, the model reconstructs the full-dose central slice. Training uses a global L1 loss for the whole image, a region of interest (ROI) based L1 loss for annotated lesion and noise crops, and a contrastive ROI loss to distinguish lesion and noise representations. Interleaved ROI-Focused Optimization (IRFO) is used during training to repeatedly revisit the limited ROI annotations. The dataset consists of 83 anonymized adult patients split into 60 training, 10 validation, and 13 test subjects. Low-count images were generated retrospectively from list-mode acquisitions, with the 25% dose level used for the training paired with corresponding full-dose slices. Under the 25% low-dose input setting, CMCL-Net improved whole-image peak signal-to-noise ratio (PSNR) by 0.188 dB and tied with the best structural similarity index measure (SSIM) value. It also improved ROI-PSNR by 0.178 dB and ROI-SSIM by 0.006 over the strongest competing baseline. Experimental evaluation and ablation studies show that multi-slice input, ROI-focused supervision, IRFO, and contrastive learning improve low-count PET restoration, especially in annotated lesion and noise regions.
DATE: June 23, Tuesday @ 09:30
Place: EA 409