Online program Overview Session: MIC-14

Abstract

Dual-flat-panel positron emission tomography (PET) systems offer strong advantages such as improved sensitivity but suffer from reconstruction artefacts due to missing projection data. We propose a deep learning model to correct these artefacts in the shearlet domain. The shearlet transform provides a multiscale, multidirectional analysis of the image, decomposing it into its shearlet coefficients, each corresponding to a subset of projections. A U-Net is used to correct for artefacts in these coefficients, but by only modifying coefficients corresponding to missing projection data, we can obtain an overall more robust algorithm. The model is trained and evaluated on data from a dual-panel PET system simulated in GATE, using real patient PET scans as the ground truth activity maps. The resulting artefact-corrected images showed good visual agreement with reference images from a system with full angular sampling. The standardized uptake value (SUV) mean of a hotspot also showed improvement (being closer to the reference value), but further investigation is required with regards to the consistency of quantitative accuracy of the model. Future work will also include an extension to 3D, to better leverage projection data along the axial dimension.

Abstract

Self-collimating (SC) SPECT employs a multi-layered, interspaced mosaic array detector structure to enable simultaneous gains in sensitivity and resolution. The complex configuration, however, demands a new set of imaging metrics and strategy to guide system design and optimization. In this work, we propose a Multi-Objective Bayesian Optimization (MOBO) framework based on four voxel-wise imaging metrics calculated from the projection probability density functions (PPDF) of the system: Angular Sampling Completeness Index (ASCI), average PPDF FWHM (full-width at half-maximum), Multiplex Index (MPXI), and sensitivity, to predict a design's performance and efficiently navigate the design space.

Through analytical simulations, we study a SC small animal SPECT that consists of a cylindrical tungsten high-resolution collimator ring surrounded by four concentric detector rings. In an initial test, we seek optimal design parameters for the aperture diameter, the number of apertures on the collimator ring, and crystal counts in two independent detector rings. The MOBO agent uses independent Gaussian process surrogates with log noisy expected hypervolume improvement to propose Pareto-improving configurations sequentially.

Results of MOBO search confirm that the four objectives genuinely conflict, which is desired, and provide recommended parameters balancing resolution, sensitivity, angular coverage, and multiplexing. Validation studies of the MOBO recommendations are underway.

Abstract

Cascade γ coincidence imaging can provide stronger geometric constraints than conventional single-photon imaging by exploiting the coincident detection of two cascade photons, but its reconstruction is challenging due to the complex system response and spatially non-uniform sensitivity. This work proposes a physics-prior-driven deep reconstruction framework for small-animal cascade γ coincidence imaging. A direct back-projection(DBP) reconstruction image was used as a physics-informed initialization, and a cascade sensitivity map was incorporated into a gated 3D ResUNet as a system-level spatial prior. Patch-based training was further adopted for sparse point-source reconstruction to reduce foreground-background imbalance. Monte Carlo simulations showed that, for a single-point source, the proposed method reduced the mean FWHM from 1.50 mm to 0.87 mm, increased the PSNR from 43.85 dB to 54.48 dB, and reduced the peak localization error from 0.25 mm to nearly 0 mm compared with DBP reconstruction. For the hotrod phantom, the foreground MAE decreased from 0.886 to 0.142, indicating improved recovery of extended active structures. The sensitivity map comparison further showed that different incorporation strategies lead to different trade-offs between foreground recovery and background suppression. These results demonstrate that the proposed framework provides a feasible route toward fast, sensitivity-aware, and physically informed reconstruction for cascade γ coincidence imaging.

Abstract

Bone metastases are common in advanced malignancies, and SPECT/CT provides complementary 3D radiotracer uptake and anatomy for screening and follow-up. Automatic whole-body lesion localization can reduce reading burden and improve consistency in patient-level assessment. However, it remains challenging because a detector needs to balance high-recall whole-body search, false-positive suppression, and fine-resolution local detail under memory and runtime constraints. To address this challenge, we propose a two-stage retrieval-augmented lesion detection framework: a 3D Swin Transformer-FPN-FCOS detector retrieves candidates from coarse-voxel-spacing whole-body SPECT/CT, and a candidate query retrieves fine-voxel-spacing local/context SPECT/CT evidence tokens for classification, box refinement, and rescoring. In validation, the base model achieves free-response receiver operating characteristic (FROC) scores of 0.511/0.471 for Mean-FROC/Low-FROC and reduces inference time from 14.9 to 3.6 s/patient; a final reranker raises Mean-FROC to 0.530 at 5.5 s/patient.

Abstract

Lung cancer remains a leading cause of cancer-related death, and distinguishing between the two major non-small cell lung cancer (NSCLC) subtypes, adenocarcinoma (ADC) and squamous cell carcinoma (SCC), is essential for treatment planning. To evaluate whether data augmentation strategies can improve subtype classification on computed tomography (CT) images with reduced annotation burden, we compared pseudo-color mapping, adaptive histogram equalization, pixel inversion, and rotation-based augmentation in combination with tumor-centric and lung-centric region extraction. Using the public NSCLC-Radiomics dataset, features were extracted with a pretrained ResNet-50 network and classified with support vector machine (SVM) and k-nearest neighbors (KNN) models, with all augmentation and feature selection performed within each cross-validation fold. Rotation consistently improved classification, with the best single model reaching 89.36% patient-level classification accuracy. Tumor-centric and lung-centric strategies showed complementary error patterns, and an ensemble of five complementary models achieved 100% patient-level classification accuracy on this cohort. The findings suggest that complementary augmentation pathways can diversify feature representations and improve classification stability in small-sample studies.