Bahram Parvin

Three-dimensional imaging of organoids typically relies on structured illumination microscopy (SIM), confocal microscopy, or classical deconvolution, but these approaches are limited by low-throughput acquisition and dependence on specialized hardware and software. Here we introduce Deconv3D, a lightweight transformer–convolution model for rapid, learning-based restoration of volumetric fluorescence images from raw z-stacks. It is trained to approximate the image quality of SIMdeconvolution workflows without requiring a point-spread-function model or specialized optics. Deconv3D uses attention-weighted fusion (AWF), which combines convolutional feature extraction with window-based self-attention. We trained the model on paired raw and SIM-deconvolved 3D images of organoids from the MCF10A progression series and MCF7 cells stained for DAPI and 𝛽-catenin, using data acquired with a high-quality objective. We then evaluated it on images acquired withlower-quality objectives and on a second microscopy platform. With only 2.23 million parameters, Deconv3D closely reproduces the SIM-deconvolution reference (PSNR 30.33 dB, SSIM 0.875). It outperforms existing CNN-, GAN-, and transformer-based restoration methods and generalizes across two distinct optical platforms and objectives. Localization-specific training further improves reconstruction accuracy, and ablation studies confirm the contribution of AWF. Deconv3D therefore provides fast, hardware-independent restoration that matches SIM-deconvolution quality, supports high-throughput imaging and downstream quantitative analysis. The code, dataset, an interactive web application, and a usage video are publicly available.

Three-dimensional imaging of organoids typically relies on structured illumination microscopy (SIM), confocal microscopy, or classical deconvolution, but these approaches are limited by low-throughput acquisition and dependence on specialized hardware and software. Here we introduce Deconv3D, a lightweight transformer–convolution model for rapid, learning-based restoration of volumetric fluorescence images from raw z-stacks. It is trained to approximate the image quality of SIM deconvolution workflows without requiring a point-spread-function model or specialized optics. Deconv3D uses attention-weighted fusion (AWF), which combines convolutional feature extraction with window-based self-attention. We trained the model on paired raw and SIM-deconvolved 3D images of organoids from the MCF10A progression series and MCF7 cells stained for DAPI and 𝛽-catenin, using data acquired with a high-quality objective. We then evaluated it on images acquired withlower-quality objectives and on a second microscopy platform. With only 2.23 million parameters, Deconv3D closely reproduces the SIM-deconvolution reference (PSNR 30.33 dB, SSIM 0.875). It outperforms existing CNN-, GAN-, and transformer-based restoration methods and generalizes across two distinct optical platforms and objectives. Localization-specific training further improves reconstruction accuracy, and ablation studies confirm the contribution of AWF. Deconv3D therefore provides fast, hardware-independent restoration that matches SIM-deconvolution quality, supportshigh-throughput imaging and downstream quantitative analysis. The code, dataset, an interactive web application, and a usage video are publicly available.

Automated 3D Radiological and Microscopy image analysis enables quantitative therapeutics assessment. However, accurate segmentation and smaller footprints are needed. To address these issues, the current study introduces the Multi-Aperture Transformer-Convolutional Network (MAFTC-Net) with a custom loss function incorporating principal curvatures to penalize topological dissimilarities. The multi-aperture modality represents each image patch as a series of images with smaller apertures along the same central field of view. Each aperture has a pyramid representation, with the advantage of maintaining the original image resolution. The MAFTC-Net encoder comprises four Swin Transformers, where a convolutional network processes the output of each transformer to accentuate local features within a global context. Next, the outputs of the transformer and convolution blocks are integrated with a 3D fusion block. MAFTC-Net was trained on the Synapse multi-organ, ACDC, and 3D Organoid datasets, achieving an average Dice score of 90.18±0.04, 93.55±0.05, and 95.13±0.03, respectively. The improved performance also comes with the added benefits of reduced complexity with 38.2 million parameters.

Experimental approach. A, The endogenous epithelium of the inguinal mammary glands was surgically removed from 3-week-old mice. These mice were then aged to 10 weeks, irradiated with 50 cGy of γ-ray, and transplanted with Trp53-null mammary fragments. Half of the mice received low-dose aspirin in their drinking water for 6 months. All mice were monitored by palpation, and subsets were euthanized at 4 and 8 months for plasma collection. The remaining mice were observed until tumor development or the study’s conclusion at 550 days after transplantation. Tissue samples were collected from mice that remained tumor-free up to 550 days of age. B, Graphical workflow of the proteomics. Asp, aspirin; Hp-RP, high pH reversed-phase.

Differential plasma protein abundance as a function of treatment and time. Unsupervised clustering of plasma protein abundance annotated for treatment group [control, aspirin (ASP), IR, and IR + aspirin] and time of post-transplantation (4, 8, and 18 months) of tumor free and TB mice. The key pathway enrichment of proteins in the four clades is indicated in the sidebar. Unsupervised clustering using Euclidian distance and Ward.D.2.

Temporal expression of proteins as a function of treatment. A–C, Significantly differentially expressed proteins (Student t test) as a function of time and treatment. D, C4b-binding protein levels across time for each treatment (sham, purple; IR, turquoise; aspirin, green; and IR + aspirin, pink). Error bars represent SE, and asterisks denote statistical significance determined by a Student t test. ASP, aspirin.

The risk of breast cancer increases decades after ionizing radiation exposure, thereby linking aging intrinsically to the evolution of cancer. We hypothesized that radiation accelerates aging and carcinogenesis through similar pathways, specifically low-grade systemic inflammation. In this study, we used the radiation-genetic mammary chimera model to examine the differential expression of 532 plasma proteins in BALB/c female mice between radiation exposure and experiment termination at 18 months. Mice were sham irradiated or irradiated with 50 cGy prior to being orthotopically transplanted with syngeneic Trp53-null mammary epithelium and half were treated for 6 months with anti-inflammatory low-dose aspirin. Plasma was collected at 4, 8, and 18 months from non–tumor-bearing mice and from those that had developed tumors between 12 and 18 months. Plasma quantitative proteomic analysis identified significant alterations in proteins involved in the inflammatory response in irradiated mice as a function of age. Levels of C4b-binding protein were decreased at 4 months in irradiated mice compared with controls, which was blocked in aspirin-treated irradiated mice. Notable differences in the expression of proteins associated with the inflammation were evident in tumor-bearing versus similarly aged mice. Complement components C1qA, C1qB, and C1qC were significantly increased in tumor-bearing mice that had been irradiated, whereas similarly aged mice without tumors displayed a decline in complement system activity. The specific changes in the complement system, which mediates adaptive immune function, following radiation exposure may contribute to cancer progression as a function of age.Significance:Women treated with radiotherapy as children or young adults bear an increased breast cancer risk, which is more likely to be aggressive, hormone receptor negative, and immune poor. Understanding radiation effects that could be modified after exposure may lead to prevention strategies. Consistent with our hypothesis that systemic inflammation contributes to risk, the plasma proteome from mice undergoing mammary carcinogenesis demonstrates changes in the complement system.

Figure S2. Supervised heatmap of differentially abundant plasma proteins across time points.